Skip to content

A Civilization-style strategy game about governing Israel from 15 May 1948. Real ministers, real map, every fact sourced. Unity + deterministic C# sim.

323 KB
Raw Blame

Rules of the GDD v0.1 MVP

The rules of the 1948–1951 MVP in docs/GDD.md, as Phase 1 built them behind the systems.loop switch (the approved plan, §6). Since the cutover (Phase 2 PR 1) the real game runs them. Balance values are the owner's.

15 October 2026 (owner, Q-CUT-1: "Labelled design values"). From the cutover, every balance number is a labelled design value in content/, tagged on its line and listed by DesignValuesTests (R-LOOP-005). This replaces controller ruling 2 of 12 October 2026, under which placeholders lived only in the dev and test overlay, never in content/.

Turns

R-TURN-001: The turn length on each security rung (1 to 5) is read from the rules key turn_lengths of the era the turn starts in. Without the key it is quarter, quarter, quarter, month, week, the legacy table. The real game sets all five to month (D-B1), so turns from 15 May 1948 end on the 15th of each month.

R-TURN-002: An event that asks for a decision opens its window on the 15th on or before the historical decision, so the card reaches the player on the turn before it (#567 §4, owner, 15 October 2026). The exceptions are listed in the tests: first_truce (+5 days) and rhodes_talks (+2 days) arrive late by design, israel_bonds opens on its historical date (10 May 1951): the owner chose "Keep it on 10 May and record this as an exception to rule R-TURN-002" (4 November 2026, #2551); hula_drainage opens on 7 March 1951, the day of General Riley's memo asking that the works stop (they began about 14 February), so it comes on 15 March: the owner chose "Record as exception (Recommended)" (6 November 2026, #2551, #2833); and the five war cards PR 1a rewrites as decisions during or after the fact.

R-TURN-003: In the historical run, no more than 6 months pass from one card to the next (or from the last card to the game's end), and no more than 2 when the country is at full war (security level 5) on any turn of the stretch (owner, 15 October 2026, Q-CUT-8: "6 months, 2 in wartime"; the M3 gap ruling restated in months).

R-TURN-004: A move (Move) and a unit order (UnitOrder) other than Send home take effect at once and are final (the owner, 23 October 2026: "No undo, like Civ"): each is committed to the turn in progress, in the order given, and the game shows the state after them (Simulation.Preview). Every other order waits for End turn, may be changed (and withdrawn where the panel offers it) until then, and costs and runs as before. End turn runs the route step (R-MOVE-009), then the committed orders in the order given, then the waiting orders in the order queued, except that dollar conversions queued after the first Arms purchase run just before it, in their queue order (R-MONEY-004, R-ARMS-002), as one batch (Simulation.Step), so a waiting order never changes the result of a committed one (spec 2026-10-23 section 3).

Then changed (4 Nov 2026): the owner chose the option labelled "Reword the rule", whose text read "Change R-TURN-004 to "may be changed (and withdrawn where the panel offers it)"; no UI work." (maintainer reading, not part of the owner's option: in today's UI some waiting orders cannot be changed once given. An event choice and a capture or occupation choice, once made, stand until End turn (their card no longer shows: EventCardPresenter.cs:226-230, CaptureDialogPresenter.cs:110-149); Reinforce, StopWork, CancelProject and Send home cannot be changed once queued (their control hides or is disabled: InfoPanelPresenter.cs:387-396, :862; GddHexPanel.cs:257-272, :341-343).)

The loop switch

R-LOOP-001: systems.loop in eras.yaml is legacy (the default) or gdd, and every era has the same value. Its ruleset id, which a save records, is legacy or gdd-0.1. Optional systems.ruleset selects the save identity explicitly: legacy requires the legacy loop, while gdd-0.1 and gdd-0.3 require the GDD loop. Absent means the original loop-based inference. Every era must resolve to the same identity. Unknown, empty or incompatible selectors are refused at load; gdd-0.3 also requires a non-null gdd03-balance.yaml catalogue. The shared Content.UsesAttention predicate describes this identity. Shipping eras still select GDD 0.1 implicitly (#1687, part 3). Since #1688 part 1 the sim reads the predicate: gdd-0.3 has no Attention (R-GOV-002).

R-LOOP-002: The GDD state (cities, units, the next unit id, completed objectives, the score) is written into the state hash only when it is set. With the legacy loop the hash text is unchanged, byte for byte.

R-LOOP-004: The GDD game ends on the rules key mvp.end, 15 January 1952 in the real content (owner, 15 October 2026, D-E5: "End on 15 Jan 1952"), after 44 monthly turns. The end is a boundary the last turn reaches, like the last era's end; the game is then over (mvp_ended) and raises no new event, except the card of an objective that turn recorded, or of an earlier one whose card has not fired yet (R-OBJ-003). It is the same in every era, after the start and before the last era's end (so the game ends as mvp_ended, not era_ended). The eras after it stay in the repository and load (G2).

R-LOOP-005: Every balance number of the GDD loop (all of cities.yaml and unit_types.yaml, the rules' mvp: block, and the GDD orders' attention_cost) carries the tag # DESIGN VALUE: owner-set balance, not a historical figure and its decision row on its own line, and DesignValuesTests lists every one of them: a number missing from the list, or a listed one without the tag, fails. For the separately identified gdd-0.3 ruleset, the design catalogue is content/gdd03-balance.yaml, transcribed from #1668 docs/gdd/balance.md at 605277df: one key per source key (including project and biome rows), with Agreed, Default or Deferred basis and provenance; mixed rows label their parts separately. DesignValuesTests checks this inventory and the basis distinctions. Deferred start.army remains symbol A: the loader refuses any numeric guess or unapproved relabelling, pending #1717 and validation of the full 65-unit partition. The catalogue does not switch either existing loop or supply production defaults; legacy/gdd-0.1 keep the preceding tagged-value contract. The rows #2115 lists, which the owner ruled on after the transcription (30 October and 1 November 2026), hold the ruling's value and their note names the ruling and its date; of them, the Open-mode start rows (D2), the transit charge and port-queue penalty (post-MVP) and the declined unpaid-upkeep deterioration are Default, not Agreed. Other post-MVP rows that the inactive fixtures read as Agreed keep that label for now (#2115). See the catalogue mapping for derived/mixed bases and the decision summaries of the six owner rulings (the rulings as posted are linked there). Under gdd-0.3, Attention is absent: no pool, price or cost-map requirement; the older tagged Attention contract remains only legacy/gdd-0.1 (#1688, part 1). For gdd-0.3, current owner: loop-switch.

The ID R-LOOP-003 (the Phase 1 overlay, MvpSlice.Overlay) is retired with the overlay at the cutover and is not reused.

Synthetic #1716 fixtures provide an internal immutable monthly snapshot for future gdd-0.3 work (R-LOOP-001/002; #1668, loop-switch.md:17-39 at 605277df). There is no shipping caller, resolver integration, persistence or activation. Consumers must revalidate eligibility; R-LOOP-004 is unchanged. The month's work and course snapshot is taken after that turn's orders (Amir's 28 October answers 4 and 5, owner-relayed: "Training: occupies people from the turn training starts until it finishes"; "Job changes: take effect immediately. assigning the same job is a no op"). A job change made this turn is in this month's snapshot, a re-assignment to the same job leaves the state and its hash unchanged, and a course started this turn shows its pupil on the course and off work in the same snapshot (R-CITY-001/002, R-TRAIN-005, R-TRAIN-008). When a course end or a cancel releases people was answered on 1 November 2026 (#1668 Q43(b)/(c)). A cancel releases them at once: this month's snapshot shows the teacher and pupil without a job, a job given that turn is in it, and either may start another course that month (R-TRAIN-007, R-TRAIN-008). Graduation releases them in the turn the last course turn counts: neither can start a course that month, and since the graduation month's snapshot still shows the course, work starts from the next month's snapshot (R-TRAIN-006, R-TRAIN-008).

Settlement control

R-CTRL-001: Only a settlement in settlement_targets.yaml may be captured. Each target is an arab_town row of map/settlements_1948.yaml (Q-TGT-1 (a), owner ruling of 18 October 2026), stands on a land hex of map/hexes.json that is its own hex in map/settlement_hexes.yaml (one named exception: Ramla's target is (-3, 24), the neighbour nearest Ramla's point, because at 4.4 km Ramla shares its hex (-4, 24) with Rehovot and a unit there would capture it; owner on #1796, 26 October 2026, "Contested + move target"; ContentLoader.TargetOnANeighbour; a retreat or a Raid's withdrawal never ends on that hex, and a depopulated Ramla closes that hex, not Rehovot's), names a sourced operation (each source an https URL with a checked date) and carries a review status; any other row is a load error naming file and field. The shipped list holds the six rows the owner approved on 19 October 2026 and filled on 21 October 2026 (#793: jaffa, acre, nazareth, lod, ramla, beersheba; Gaza excluded), which the owner sensitivity-reviewed on #793. #799 adds each row's residents outcome (R-CTRL-016) and returns every row to status: draft until the owner reviews those fields, so the release gate holds them (D-F5).

R-CTRL-002: Settlement control is an ordinal SortedDictionary<string, SettlementControl>. It is written to the state hash as ctrl{...} only when non-empty, in ordinal id order; each record field is written only when it differs from its default (R-LOOP-002). An ownerless, non-depopulated record is the lasting marker that a captured place returned to enemy control; it retains the capture date and whether the residents fled (owner answer, 4 November 2026, #2542: "Keep a 'lost' marker (Recommended)"). This representation and the rules below are shared by gdd-0.1 and gdd-0.3; the legacy loop remains unchanged. On the map, a settlement-control record overlays only that settlement's displayed hex on the dated historical control layer: an Israeli owner uses the Israeli colour and an ownerless record the enemy colour; fronts, zones and armistice lines remain dated display data. The selected-hex panel shows the live holder first and the settlement state detail second (coordinator decision on #2542); with no live record, both map and panel keep the dated path unchanged.

R-CTRL-003: Residents and combatants are distinct fields. Residents are a state, inhabited, fled or depopulated, not a count (Q-CAP-2 (b)). Fled is the record's Fled, set at capture from the target row's residents: fled and never cleared (OWNER QUESTION 2 (b), owner, 24 October 2026, #799); flight is not Deport: it adds nothing to DisplacedByPlayer and costs nothing. No rule reads residents as combatants, prisoners, refugee units, labour or a resource; annexation is the one place a later owner-approved rule may say otherwise.

R-CTRL-004: The capture check runs after the route step and each order, and still in resolver step 3 after enemy combat, in ordinal settlement id (Q-CAP-1 (a), Q-CAP-3 (a), owner rulings of 18 October 2026). A target that starts under enemy control, one listed in mvp.enemy_held_localities (owner, 21 October 2026, #796), is captured at once when Israeli military units are on its hex and no enemy military unit is; only enemy military units are combatants (Q-CAP-3 (a)), and a halutz holds nothing (#796). Passing through a target does not capture it; only a move's or go-to's stopping hex counts (owner, 23 October 2026, OQ-2: "Instant"). Its record names israel as owner and has the state awaiting_choice; the preview leaves Since unset, and Step stamps the turn's end before the resolver, preserving the date through a same-turn answer. The answer stays deferred until End turn (R-TURN-004). A capture remains when the unit leaves; the player's units stay on the hex. A captured target still awaiting the choice that only enemy military units are on is under enemy control again, and its record becomes the ownerless marker of R-CTRL-002. A target not enemy-held at the start (Jaffa, owner ruling of 18 October 2026 on #793, "Only in a what-if") is never captured. The check is an integer rule on unit positions and draws nothing from the RNG. Enemy captures of the player's localities (Q-TGT-2 (c)) are R-CTRL-006. Its old blocker (#796), a Syrian unit starting on Degania Alef's hex, is gone since the 4.4 km grid (#1796): the Syrian start is (13, 10) and Degania Alef is (12, 11). Every ownerless, non-depopulated record is no record for the capture check, so the place can be captured again; the new record keeps its fled value and takes the new turn's date. Stamp dates every new settlement record, including the stayed lost marker, as it dates Deport's (OWNER QUESTION 2 (b), owner, 24 October 2026, #799).

R-CTRL-006: Enemy units take the player's localities (Q-TGT-2 (c), owner, 18 October 2026, https://github.com/eyalgolan/statecraft-israel/issues/783#issuecomment-6249796458: "c + in what if scenarios that are outside of the historical timeline we need to research cards"; option (c) is option (b), "enemy units capture them, with a state change and no outcome choice", with sourced cards for places sources tie to a dated loss). A player's locality is every settlements_1948.yaml row that is not an arab_town and not on mvp.enemy_held_localities (owner, 21 October 2026, https://github.com/eyalgolan/statecraft-israel/issues/796#issuecomment-6315647311: "Default 3 as proposed (the player's localities)"; the proposal: "An enemy military unit holding one with no Israeli combatant changes its state, with no choice. Israeli units holding it again change it back."). In resolver step 3, after R-CTRL-004's check, each player's locality is checked in ordinal id. One with no record that only enemy military units are on (a halutz holds nothing, #796) is lost: it gets R-CTRL-002's ownerless marker, dated the turn's end, and one place_lost line, "{name} is lost: enemy units hold it and no Israeli unit is there". A lost one that only Israeli military units are on loses that marker again and gets one place_retaken line, "{name} is retaken: Israeli units hold it again and no enemy unit is there". The check is an integer rule on unit positions and draws nothing from the RNG. As R-CTRL-005, it writes the settlement state only: it offers no choice and changes no resident, unit, resource or score. An annexed target is taken the same way, on its target's hex (R-CTRL-001; Ramla's is (-3, 24)) (the owner, 8 November 2026, answering the questions on #2955: "Can be taken", the option "Enemy units can take an annexed town like any Israeli town, and Israeli units retaking it make it annexed again"): the record loses only its owner, so it is enemy-held again, with the same place_lost line; Israeli military units alone on it give the owner back, with the same place_retaken line, so it is annexed again with the record it had. A retake is no new capture (R-CTRL-004 skips an ownerless annexed record), and nothing R-CTRL-016 added is taken away or added a second time. An enemy move ends on a player's locality or an annexed target Israel holds only when that hex is where the move is sent, and no enemy path passes through one (R-ENEMY-001; the owner, 8 November 2026, answering the questions on #2955: "Destination only (Recommended)"), so a march never takes a place on its way. No scripted enemy anchor is on a player's locality, so the shipped goldens lose none. On a city's hex the war-loss check (R-END-003) still reads the enemy unit as before. The sourced loss cards and the what-if cards of Q-TGT-2 (c) are a later content PR (#1305).

R-CTRL-005: Capture writes the settlement state only. It changes no resident field except that a target whose row says residents: fled is recorded with fled=1 and its capture line says its residents fled (OWNER QUESTION 2 (b), owner, 24 October 2026, #799), no unit, no resource, parameter, city or score, and counts towards no objective (constraint 6 of the Phase 6 plan; impact plan section 4.3).

R-CTRL-007: ResolveCapture(settlement, outcome) answers a capture. It is one command kind for every outcome, with one attention_cost.resolve_capture (Q-TECH-3, owner, 18 October 2026, #783 batch 1), set at 0 as a design value proposed on #797 (D-A6: unit orders are free). It is valid only for a settlement whose record is awaiting the choice, including the turn of its instant capture until it is answered, and only for an outcome that is built. The answer remains deferred and runs at End turn (owner, 23 October 2026; R-TURN-004). Otherwise it is refused with a reason, the state unchanged and the Attention refunded: the place is not awaiting a decision, the outcome is not known, or Deport is not in play: with mvp.deport_in_play off, as in the real content until the owner has reviewed Deport's cards and dialog (Phase 6 PRs 9 and 10), it is refused as "not built yet" (R-CTRL-018). The legacy loop refuses it before charging Attention.

R-CTRL-008: SettlementControl.CaptureOptions(content, state, settlement) lists the outcomes in one fixed order, Raid and withdraw, Occupy, Deport, so Deport is never first (impact plan section 4.1 item 1). An option carries no default or selected flag. Each option has either the reason it is refused or a preview. A refusal carries a stable id from a fixed list and the values its line needs beyond the place itself (for Attention, what the option needs and what is left), so the dialog shows a reviewed line in the player's language and never reads the English reason the command returns (Q-TECH-5, #802). A preview is the change resolving it makes at once to Attention, Money (the Treasury and the debt), Food, Arms, Social cohesion, International Support and the score, computed by resolving it on a copy of the state, so the preview is the outcome (spec section 2.6). A place not awaiting the choice offers no option. A capture the player does not answer waits, and the units stay (Q-HIST-1 (a), owner, 18 October 2026, #783 batch 3). On the turn that reaches the game's end, where resolver step 10 sits and before objectives and score, every capture still awaiting the choice counts as Raid, in ordinal id order (owner, 18 October 2026 on #797, "Counts as Raid"): the units leave as R-CTRL-010 says, nothing changes for the residents and the score gains nothing. A unit with nowhere to go stays while the others leave, and the ownerless marker remains all the same. No unanswered choice ever resolves as Deport.

R-CTRL-009: HistoricalAI answers every capture awaiting the choice with Raid, in ordinal id order, leaving out a Raid the sim would refuse (a unit with nowhere to go), which then waits, and the reference player sends HistoricalAI's orders, so it does too (Q-REF-1 (a), owner, 18 October 2026, #783 batch 3). Neither ever issues Deport, and nothing delegates a capture: EventSystem.Delegate answers events only (impact plan section 4.1 item 1).

R-CTRL-010: Raid and withdraw (Q-RAID-1 (a), owner, 18 October 2026, #783 batch 3). The player's units on the place's hex, soldiers and halutzim alike, leave in ascending id, each to the first neighbouring hex in the fixed neighbour order (R-MOVE-002) that is land, holds no enemy unit, is not an enemy-held locality (R-MOVE-005) and has room under the stack limit (R-MOVE-003). Leaving costs no movement points, and the units keep their orders, as after a Move that ends on another hex: a fortify is lost at the turn's order step (R-ORDER-002); both are defaults proposed on #797. The record becomes ownerless, preserving Since and Fled, so the place is under enemy control again at once and its residents are unchanged. Raid changes nothing else: no resource, parameter, city, score or objective, and no RNG draw. If a unit has no such hex, Raid is refused with a reason and the state is unchanged, so the choice waits. For a fled place the line says "Its residents fled when it was captured" instead of "Its residents are unchanged" (OWNER QUESTION 2 (b), owner, 24 October 2026, #799).

R-CTRL-011: Occupy (Q-OCC-1 (a), owner, 18 October 2026, #783 batch 3): ResolveCapture(settlement, occupy) keeps the place under the player's control. It needs an Israeli military unit on the place's hex, the one unit that stays (a halutz holds nothing, #796), and is refused with a reason otherwise, the state unchanged. It moves no unit. It changes the record's state to occupied and starts its timer (R-CTRL-012), and nothing else: no resident field (the place stays inhabited, R-CTRL-003), unit, resource, parameter, city, score or objective, so its preview (R-CTRL-008) is all zeros. Each turn's step 10, while the place is occupied or awaiting EndOccupation, the Treasury pays mvp.occupation_upkeep_per_turn for it, borrowing what it lacks (R-MONEY-003); the Occupy turn pays too. The upkeep is a design value accepted by the owner on 22 October 2026 (https://github.com/eyalgolan/statecraft-israel/pull/1324#issuecomment-6338594839): IL 0.005m, one month of income_per_job_per_year for the one population unit that stays, derived as held_soldier_cost_per_turn is. It comes on top of the garrison formation's own military_upkeep_per_turn (IL 0.05m, R-UNIT-010), so holding a place costs IL 0.055m a turn in all (owner ruling, 22 October 2026, https://github.com/eyalgolan/statecraft-israel/pull/1324#issuecomment-6341877796). Money is shown to one decimal of IL millions, so when the Treasury has to borrow for it the Why line shows the borrowed amount, at most IL 0.005m, as "IL 0.0m". For a fled place Occupy's line ends "Its residents fled when it was captured", and Occupy and the timer keep fled=1 on every record (OWNER QUESTION 2 (b), owner, 24 October 2026, #799). The owner's rulings of 27 October 2026 on #1712 (https://github.com/eyalgolan/statecraft-israel/issues/1712#issuecomment-6461246243) keep this flow for capture, "Existing Raid/Occupy", and the occupation "3 turns, keep upkeep" (R-CTRL-012); the owner's ruling of 1 November 2026, "30 Oct wins for the MVP (Recommended)", leaves Occupy as written here for the MVP. It holds unchanged under gdd-0.3: a capture, Occupy, a withdrawal and a retake end every turn as under gdd-0.1, the Attention pool aside (OccupyTests.WithdrawingOnTurnTwoAndRetakingIsTheSameUnderGdd03).

R-CTRL-012: The occupation timer (Q-OCC-3 (a), owner, 18 October 2026, #783 batch 4: one X for every place). Occupy sets the record's Turns to mvp.occupation_turns, X, a design value: 3 monthly turns (owner, 27 October 2026, #1712, "3 turns, keep upkeep", https://github.com/eyalgolan/statecraft-israel/issues/1712#issuecomment-6461246243; confirmed by the owner on 1 November 2026, "3 turns (27 Oct)"), replacing the 6 monthly turns the owner accepted on 22 October 2026 (https://github.com/eyalgolan/statecraft-israel/pull/1324#issuecomment-6338594839). Resolver step 10, after step 9's arrivals and before objectives and score (step 11), takes each occupied place in ordinal id order and counts one turn off, the Occupy turn included; at 0 the place's state is awaiting_end, where it awaits EndOccupation (Phase 6 PR 7) and the count stops. Turns is written to the hash as turns= only when non-zero (R-CTRL-002). The step draws nothing from the RNG.

R-CTRL-013: Withdrawal before the timer ends (Q-OCC-2 (a), owner, 18 October 2026, #783 batch 3: allowed, and it becomes a Raid). An occupation lasts while an Israeli military unit is on the place's hex. When step 10 finds none on an occupied or awaiting_end place (the player moved them out, or they were lost), the occupation ends as a Raid does: the record becomes ownerless, preserving Since and Fled, so the place is under enemy control again at once, its residents unchanged, and no upkeep is paid that turn. Nothing else changes, and nothing is refunded. A halutz left on the hex keeps nothing. For a fled place the line says the residents fled when it was captured (OWNER QUESTION 2 (b), owner, 24 October 2026, #799). The owner's rulings of 27 October 2026 on #1712 ("Existing Raid/Occupy", "3 turns, keep upkeep"; https://github.com/eyalgolan/statecraft-israel/issues/1712#issuecomment-6461246243) and of 1 November 2026 ("30 Oct wins for the MVP (Recommended)") leave it as written: a withdrawal on the occupation's second turn returns the town and clears its timer, and a retake is a new capture, whose Occupy starts a fresh timer. It holds unchanged under gdd-0.3 (OccupyTests.WithdrawingOnTurnTwoAndRetakingIsTheSameUnderGdd03). An operation card's order never takes the last Israeli unit off an occupied place (R-ENEMY-005; owner, 2 November 2026, #2087: "Keep a garrison in Nazareth (Recommended)"). That rule covers only the cards' orders: the player's own orders (a move, a go-to, Send home) and combat (a unit's loss, its surrender, or its retreat after an enemy attack, R-COMBAT-003) can still leave an occupied place with no Israeli unit, and step 10 then ends the occupation as above.

R-CTRL-014: EndOccupation(settlement, outcome) answers the end of an occupation (#799). It is one deferred command kind (not instant, R-TURN-004) with one attention_cost.end_occupation, set at 0 as a design value proposed on #799 (D-A6: unit orders are free). It is valid only for a place in awaiting_end with an Israeli military unit on its hex, as Occupy needs (R-CTRL-011), and an outcome of annex or withdraw; the unit is checked for both outcomes. Otherwise it is refused with a reason, the state unchanged and the Attention refunded. It draws nothing from the RNG. Saved as endocc <settlement> <outcome>; scenario kind end_occupation. The owner's rulings of 27 October 2026 on #1712 (https://github.com/eyalgolan/statecraft-israel/issues/1712#issuecomment-6461246243) set the occupation it ends at "3 turns, keep upkeep" (R-CTRL-012), and the owner's ruling of 1 November 2026, "30 Oct wins for the MVP (Recommended)", keeps its outcomes to Annex and Withdraw for the MVP (R-CTRL-015, R-CTRL-016). Each option's preview equals its result (EndOccupationTests.TheEndPreviewIsTheOutcome), and annexing a stayed town adds one city and no unit (EndOccupationTests.AnnexingAStayedRealTownAddsOneCityAndNoUnit). It holds unchanged under gdd-0.3: Occupy, the three turns, Annex and the unit's leaving end every turn as under gdd-0.1, the Attention pool aside (EndOccupationTests.OccupyThenAnnexIsTheSameUnderGdd03).

R-CTRL-015: No puppet regime (Q-PUP-1 (c), owner, 18 October 2026, #783). SettlementControl.EndOptions(content, state, settlement) offers Annex and Withdraw, in that order (the ruling's words, "annex or withdraw"), with no default or selected flag, each with its refusal or its preview as CaptureOptions (R-CTRL-008); it is empty when the place is not awaiting the end. Withdraw is R-CTRL-010's withdrawal, the same code path.

R-CTRL-016: Annexation and the residents (Q-ANNEX-1, owner, 18 October 2026, #783; OWNER QUESTIONS 1 (a) and 3 (a), owner, 24 October 2026, #799). The record becomes owner=israel,state=annexed, keeping its capture date and fled, with no timer. A fled place adds nothing else: no city, people, housing, jobs, Food, Money, Arms, indicator change, score or objective progress ("Territory only, no gain"). A stayed place adds one city by the starting-city rule: its row's annexation.people (a sourced count the owner approves) in population units (R-CITY-007), housing equal to them, job slots equal to the employed (people times annexation.employed_share, the national share 0.439, to the nearest unit), the condition mvp.housing_start_condition, and no refugees, recruits or craftsmen, and the same approved people join GameState.Population (R-CITY-008; owner, 4 November 2026, #2608, the option labelled "Yes, add them (Recommended)"), in a group of their own, annexed, unless an initialized gdd-0.3 ledger keeps the record (the separate group and that exception are this change's design, not the owner's ruling); refused, the state unchanged, if the id is already a city. Each target row says residents: fled or stayed with at least one source (residents_sources); only a stayed row has an annexation block, whose people must be a sourced whole number; any other row is a load error naming file and field. The town stays Israeli unless enemy units take it: step 10 ignores an annexed record, so the upkeep stops and no unit has to stay, and an annexed target Israel holds is no enemy-held locality for movement (R-MOVE-005, R-ENEMY-001). Enemy units can take it like any Israeli town, and Israeli units retaking it make it annexed again (R-CTRL-006; the owner, 8 November 2026, answering the questions on #2955: "Can be taken"); while it is taken its record has no owner and it is enemy-held again. A city that exists only in state has no training facility.

R-CTRL-017: A depopulated place (Deport) is never offered or given EndOccupation, so it is never annexed (Q-ANNEX-2 (a), owner, 18 October 2026, #783). The owner's rulings of 27 October 2026 on #1712 (https://github.com/eyalgolan/statecraft-israel/issues/1712#issuecomment-6461246243) and of 1 November 2026 ("30 Oct wins for the MVP (Recommended)") leave it as written. It holds unchanged under gdd-0.3 (EndOccupationTests.ADepopulatedPlaceIsNeverAnnexed).

R-CTRL-018: Deport (impact plan section 4.1; owner ruling 1, "honest cost, no reward"; Q-DEP-1 (a), Q-DEP-5 (c) and Q-DEP-7 (c), which includes (a), owner, 18 October 2026, #783 "Accept all 16 drafts"): ResolveCapture(settlement, deport), only while mvp.deport_in_play is on (R-CTRL-007). The player's units on the place's hex leave exactly as in Raid (R-CTRL-010), and Deport is refused the same way, the state unchanged, when one has nowhere to go: it gives no military advantage over Raid and pays no occupation upkeep. Social cohesion (morale) falls at once by mvp.deport_cohesion_cost and International Support (intl_standing) by mvp.deport_support_cost, through the permanent event shift. The Treasury pays mvp.deport_money_cost at once, borrowing what it lacks (R-MONEY-003), so Deport never pays on Money (owner decision, 22 October 2026, https://github.com/eyalgolan/statecraft-israel/pull/1327#issuecomment-6337946480): a design value set at or above the most Occupy can pay in upkeep over the whole game plus the dearest city project, IL 6.25m. The place is added to DisplacedByPlayer. From then on, every turn's resolver step 10, after the occupation timers, International Support falls by mvp.deport_support_drag_per_turn for each place in DisplacedByPlayer, the Deport turn included, for the rest of the game; and International Support's ceiling is 1 less mvp.deport_support_ceiling_drop for each such place (never below 0), which every write to it respects. The four indicator values (0.15, 0.15, 0.005 and 0.25, the drag and the ceiling drop per place) and the IL 6.25m were proposed on #800 and #1327 and accepted by the owner on 22 October 2026 (https://github.com/eyalgolan/statecraft-israel/pull/1327#issuecomment-6350226175); they stay design values the balance pass (#813) may revisit. Nothing draws from the RNG. For a place whose record is fled (its residents fled at capture, R-CTRL-005) Deport stays listed third and is refused, the state unchanged, with the reason "Its residents fled when it was captured, so there is no one to deport", which the options helper shows on its button (OWNER QUESTION 4 (a), owner, 24 October 2026, "Hide it, button explains (Recommended)", https://github.com/eyalgolan/statecraft-israel/issues/799#issuecomment-6375591758).

R-CTRL-019: The residents (Q-DEP-2 (a), every resident the state holds; Q-DEP-3 (a), a count only, with no destination on the board): a Deport adds 1 to DisplacedByPlayer, a count of places, not people (Q-CAP-2 (b)), written to the hash as displaced= only when non-zero. They are outside the player's control: never refugee units, civilians, cohorts, units, prisoners, labour, housing occupants or score.

R-CTRL-020: A depopulated place (impact plan section 4.1 item 4; Q-DEP-7 (c); Q-SENS-2 (c)). Deport leaves the place's record with its capture date and depopulated=1 only: no owner, state or timer. It stays on mvp.enemy_held_localities, so the player's units cannot enter it (R-MOVE-005) and it is never captured again, as a raided place is for the enemy. It adds no housing, jobs, Food, Money, Arms or score, and counts towards no objective. With no timer, resolver step 10 neither counts it down nor charges upkeep for it. No city project may be started on its hex (refused with the reason kind depopulated), and no corridor segment through it (this stops new starts; one already under way there is not cancelled); no designated site can be on a 1948 settlement's hex or a settlement target's hex at all (#2804), and no event effect can name a place.

R-CTRL-021: Buildings (D-F3; Q-DEP-4 (a)): deportation and razing are separate rules, and there is no razing option in the MVP. The record's buildings stay at their default, and a capture offers only Raid, Occupy and Deport.

R-CTRL-022: DeportNeverDominates (impact plan section 4.1 item 5; Q-TECH-4 as drafted; Q-DEP-7 (b); Q-DEP-5 (c), "must hold with a margin"). From any state where Deport is offered, the state is cloned for Raid, Occupy and Deport, each clone resolves its outcome and is stepped by the reference policy to the game's end. At the turn's end and at the end, Deport is no better than Raid or Occupy on any measure, with no exception (owner decision, 22 October 2026, #1327): Money, foreign currency, Food, Arms, Attention, coalition stability, any indicator, the score, the objectives and the day each was met, the player's units and their strength, housing, jobs, people, built sites and the war loss, and it is below both on Social cohesion and International Support by a margin. Where a city shares the place's hex (as Tel Aviv shares Jaffa's), the city cannot build there after a Deport, so it may spend less; mvp.deport_money_cost covers that too, and Money is compared there as everywhere. Until a capture is reachable in the real content (#1304) the fixture's capture states carry the test.

Saves

R-SAVE-001: A save names the ruleset it was played under (ruleset legacy or ruleset gdd-0.1, save schema 2). A game refuses a save made under another ruleset. A schema 1 save has no ruleset line and is legacy, so the GDD game refuses it as made under earlier rules. The refusal carries a code (older_rules) the UI translates, and the sim's reason does not itself ask for a new game: the player reads that once (N4). GDD 0.3 has its own ruleset, gdd-0.3, first saved as schema 4 (#1687; owner ruling 26 October 2026: GDD 0.3 is built behind its own ruleset), while legacy and gdd-0.1 saves were schema 3; every ruleset wrote schema 5 (R-SAVE-005) and now writes schema 6 (R-SAVE-006). A gdd-0.3 game refuses a legacy or gdd-0.1 save as made under earlier rules (older_rules), before any replay. No content selects gdd-0.3 yet, so the live game is unchanged.

R-SAVE-002: A save that names a ruleset this build does not know, or a newer save schema, was made by a newer version of the game and is refused with the code newer_version (N2). A legacy or gdd-0.1 game refuses a gdd-0.3 save the same way, A gdd-0.3 save under schema 2 or 3, and a schema 4 save naming legacy or gdd-0.1, were never written by any build, so they are refused as damaged, with no code (#1687, #2605).

R-SAVE-003: Save schema 3 adds the open turn: after the last turn, a line open and the turn in progress's committed orders (R-TURN-004), one a line (a waiting order there is refused: the save was edited); the end hash is still the state's after the last finished turn. Loading replays the turns, then commits the open orders again; one refused on the same content means the save is damaged (SaveException), and on other content the rest are dropped with the replay_differs warning. A schema 2 save has no open turn and loads as before (but see R-SAVE-004). A save older than schema 5 is R-SAVE-005's, and a schema 5 one R-SAVE-006's.

R-SAVE-004: A schema 2 save of the GDD game with a player attack (order <unit> attack <q> <r>) is refused with the code older_rules (R-SAVE-001): its attacks were standing orders, so it would replay differently (the owner, 23 October 2026: "Refuse with a clear message"). So is a schema 2 GDD save whose replay under the current rules does not end on its saved hash (instant capture, the full-points step). Other schema 2 saves load.

R-SAVE-005: Since #2605 the state records the soft events the cap left waiting (R-EVT-001), so every ruleset writes schema 5. An older save of a valid schema for its ruleset (1 to 3 for legacy, 2 to 3 for gdd-0.1, 4 for gdd-0.3; schema 1 has no ruleset line, so it is legacy) still loads, and is written back in the current schema (R-SAVE-006). #2536's exception for a finished game without its ending cards applies to such a save only, as does #2542's for a same-content save without its stayed lost markers; a same-content save without both is the same game when taking both out of the replay gives its saved hash, and loads with no warning (#2733). It was played under the earlier cap, so with the same content a replay that does not end on its saved hash, or an open-turn order the replay now refuses, is the rules' change, not damage: it loads with the replay_differs warning. R-SAVE-004's schema 2 refusals come first. A schema 5 save is R-SAVE-006's.

R-SAVE-006: Since #2544 a player's order to a unit in an operation cancels the operation's remaining orders for its other units (R-ENEMY-009), so every ruleset writes schema 6. A schema 5 save, of any ruleset, still loads and is written back as schema 6. It was played before that cancel, so with the same content a replay that does not end on its saved hash, or an open-turn order the replay now refuses, is the rules' change, not damage: it loads with the replay_differs warning. The exceptions for a missing ending card (#2536) and a missing lost marker (#2542) still apply only to saves older than schema 5, as before. A schema 6 save that does either on the same content is damaged (SaveException).

Cities

Each GDD turn resolves in the order of the approved plan's section 2.3 (decision D-A6): production, then housing and Social cohesion, then arrivals.

R-CITY-001: Under legacy/gdd-0.1, AssignWorkers(city, role, n) sets how many of the city's people work in role (only job exists: a craftsman fills an ordinary job slot, R-CRAFT-002). It is refused, with the state unchanged and the Attention refunded, when the city or role is unknown, n is negative, or n is above the city's jobs or its civilians and craftsmen. An accepted order turns the city's job default off (R-CITY-009). The legacy loop refuses it before charging Attention. The gdd-0.3 population foundation uses one typed assignment per ledger id (R-POP-002); initialized ledger fixtures refuse this aggregate command path, pending producer integration (on a Historical ledger the workshop role excepted, R-ARMS-005; #1725 PR 7). Under gdd-0.3 the role workshop staffs the city's arms workshops instead (R-ARMS-005), and a job count is refused above the city's civilians and craftsmen less its staffed workshops and its builders (R-CITY-011), whose workers hold no job; the job default (R-CITY-009) fills to the same number. Under gdd-0.3 the role build puts builders on the city's projects being built (R-CITY-011).

R-CITY-002: Each turn every city's employed units add employed * income_per_job_per_year * the turn's share of a year to the Treasury (an employed craftsman earns more, R-CRAFT-002). The gdd-0.3 catalogue's industry.income tiers (ordinary or novice / apprentice / expert: 1 / 2 / 3) are multiples of this per-unit rate (owner, 1 November 2026: "Main's base x 0.3's tiers (Recommended)"; #2112). The inactive synthetic fixtures (Gdd03AssignmentFixture.IndustryRate, Gdd03CraftArithmetic) quote the catalogue multiple itself, not Money; Gdd03IndustryRates.MoneyPerTurn converts it, and no live system reads either. A staffed arms workshop's worker is not employed and earns nothing here (R-ARMS-005).

GDD 0.3 fixture clarification for R-CITY-001/002: the owner’s 28 October Q34 answer makes reassignment to the same job a no-op (owner-relayed answer). The detached generic Work fixture returns NoOp and preserves the complete state when given its current target; its different-target transition already applies immediately. This records the settled fixture result, without a production command or a new course, service or save policy.

Synthetic regional fixtures additionally enforce R-CITY-001/002 and R-CITY-006: Industry and Workshop occupancy is counted separately in each worker’s initialized region and cannot exceed its effective PlayerRegionPool job slots, including facility contributions. Reassignment excludes the mover’s current reservation before checking the destination; refusals preserve the complete state. Farm and Builder retain their explicit slot bounds. This adds no real job/facility mapping, command, production payment or gameplay activation; fixtures without regional pools keep their existing slot-only behavior.

R-CITY-003: Arrivals from the CBS series (aliyah.yaml) become refugees in the cities. Let P(t) be the people the series brings from city_arrivals_from to the date t (0 before it), summed over the origin groups in ordinal order. city_arrivals_from is 8 November 1948, the census day the cities' start values count: the census already holds the people who arrived from 15 May, so the cities take none before it (Q-ARRV-3, owner, 15 October 2026). The cities' arrived units by t are U(t) = floor(P(t) / people_per_unit), and a turn from date d to date e adds U(e) - U(d) refugees over all cities. people_per_unit and city_arrivals_from are the same in every era (the loader refuses an era that changes either), so over any run of turns the refugees added are U(end) - U(start): no unit is lost or counted twice. The units are split in proportion to the cities' census people (Q-ARRV-1): the first U units are handed out one at a time, unit n to the city with the largest remainder n * people / total people - units it holds, a tie to the lower city id (ordinal), and a turn gives each city its share of U(e) minus its share of U(d). Building the split up one unit at a time keeps it cumulative: a fresh largest-remainder split of each total could take a unit back from a city as the total grows (the Alabama paradox; on the twelve cities it first happens at 7 units). The loader refuses a GDD game with no city. With an initialized Historical ledger (the dark start, R-POP-002; #1725 PR 7) each arrived unit is one refugee record instead of a count (GDD Q11: arrivals "take new ids as refugees in a city"), registered with source key cbs/<city id> and, as its ordinal, the unit's place in that city's cumulative share from city_arrivals_from, so the same units get the same keys whichever turn brings them and a replay of a turn adds nobody twice (R-POP-001). The record is located in its city, in the region that lists it, with reception 0 and no assignment, profession or tier. The split and the city messages are as above; the aggregate city fields are not written, and PopulationSystem adds no second count of these people (R-CITY-008). A synthetic or Open ledger is still refused.

R-CITY-008: Under legacy/gdd-0.1, GameState.Population is the one record of the nation's people: it takes every arrival from 15 May 1948 and is the count the CBS anchors check. The cities are the part of it placed in towns, so their people (civilians, craftsmen, refugees, recruits and the groups in training, times people_per_unit) never exceed it (Q-ARRV-2, owner, 15 October 2026). No person is counted twice. A stayed town's annexation (R-CTRL-016) adds its approved people to GameState.Population (owner, 4 November 2026, #2608, the option labelled "Yes, add them (Recommended)"; EndOccupationTests.AnnexingAStayedPlaceAddsItsPeopleToThePopulationAndAFledOneAddsNone). They go in a group of their own, annexed, not an aliyah origin, and an initialized gdd-0.3 ledger fixture is left alone (EndOccupationTests.AnnexingIntoAnInitializedGdd03LedgerLeavesThePopulationAlone); both are this change's design, not the owner's ruling. The total (GameState.TotalPopulation) counts them wherever the game reads it (owner, 4 November 2026, the option labelled "Count them everywhere", https://github.com/eyalgolan/statecraft-israel/pull/2620#pullrequestreview-5777301237), so it is the whole population, no longer the Jewish population alone. In initialized gdd-0.3 fixtures, the own-population ledger is authoritative instead: profession, employment, training and captivity are attributes/statuses of the same ids (R-POP-001), never extra people. No old headcount is rounded into ledger records. On an initialized Historical ledger (the dark start, R-POP-002; #1725 PR 7) PopulationSystem.Apply adds no second count of the CBS arrivals, which are R-CITY-003's refugee records, and leaves GameState.Population alone; it adds the natural increase at the national rate (GDD Q11, 0.0169 a year) as records, under four rulings (owner, 8 November 2026): "Carry the fraction (Recommended)": the part of a record a turn owes is kept from turn to turn, and a record is added each time the carried sum reaches one. "Same split as arrivals (Recommended)": the new records are split over the cities by R-CITY-003's largest-remainder split of all natural-increase records so far. "Free civilian (Recommended)": each new record is a free civilian of its city, with no work experience and no reception wait. "Everyone alive (Recommended)": the base is every living record of the ledger, soldiers included. A turn owes the carry plus the rate times the turn's share of a year times that base, in the aggregate natural increase's double arithmetic; a record's source key is increase/<city id> with its place in that city's share as the ordinal. The carry and the count of these records are in the gdd-0.3 ledger's canonical text (R-POP-003) once natural increase has run, the carry as the 16 hex digits of its IEEE bits and omitted when zero (docs/determinism.md), so no earlier ledger and no legacy or gdd-0.1 game hashes differently. A negative rate has no ruling and is refused. Saving and reloading a ledger game is still held (the note before R-POP-001: "command/save/reload integration remain held").

R-CITY-004: Each turn, on the housing step, a city's refugees become civilians without a job (R-CITY-009's default may then give them jobs), floor(pool x m) of them at the Immigration minister's multiplier since the ministers (R-MIN-002; all of them before); the rest wait as refugees for the next turns. There is no hard cap (D-D1, Q-HOUSE-1, owner, 15 October 2026): the people beyond the city's housing stay in the city as its overflow (R-HOUSE-002), and residents are housed automatically (Q-HOUSE-4).

For R-CITY-004, the preceding live absorption rule remains unchanged. The inactive gdd-0.3 addition is Gdd03ReceptionFixture.ResolveWholeTurn, an internal synthetic ordinary-bed fixture requiring exact gdd-0.3 content/state and an initialized population ledger. It reads positive Agreed integer reception.capacity and reception.turns without fallback (currently 3 beds and 6 turns); the ledger remains the person/progress authority under R-POP-002/003, reused here without a new persistence contract (sim/Statecraft.Sim/Gdd03ReceptionFixture.cs:36-65@a30013863; sim/Statecraft.Sim.Tests/Gdd03ReceptionFixtureTests.cs:4-16@a30013863, :149-190; content/gdd03-balance.yaml:214-221@a30013863). Source: docs/gdd/training.md:23@605277dfa511dc512562ececeb273b8393b21d35, "Completion changes the same unit into a free civilian, retaining id and experience."; docs/gdd/balance.md:97-98 at that revision supplies the capacity and duration.

Fixture conventions only: retained opening occupants have priority and gain one tick; reaching the requirement makes the same id an unassigned free civilian, preserving history, equipment, home and location. New claims are allocated at step end by ascending person id then ordinal camp target, only at an explicitly supplied operating, exactly colocated camp with space. Admission, including reuse of a just-released bed, gives no tick. Work eligibility is captured at opening, so completion grants no work in that step; the standalone eligibility predicate requires completed reception, free-civilian status and no assignment, and does not wire job/farm/build/course entry points. Invalid retained claims (including missing/inoperative/mismatched/overfull camps or completed/non-refugee occupants), duplicate camp targets and Course/Teaching assignments are refused rather than choosing interruption or shared-bed policy. With no available camp, an unassigned refugee stays unassigned. Changes are made on a reconciled clone, with no turn/date, RNG or resource advancement (sim/Statecraft.Sim/Gdd03ReceptionFixture.cs:41-118@a30013863; behavior fixtures in sim/Statecraft.Sim.Tests/Gdd03ReceptionFixtureTests.cs:35-146@a30013863, invalid completed claims at :194-201). These ordering and timing conventions are not adopted live timing rulings.

The #1696 status comment of 27 October 2026 says, "It does not yet provide the live port queue or integrated camp routing." Live arrivals (#1695), map/region binding and approved source rows/coordinates/hexes (#1691; #1690 closed on 2 November 2026, its post-MVP regional pools moved to #2233), shared course beds (#1697), military credit/capping/deduplication and return queues (#1699), transit charging and queue-cohesion settlement (#1702/#1696), monthly timing and committing once (#1716), live activity guards, commands and save/replay integration remain pending. The fixture is "not a monthly settlement receipt or a replayable command" (sim/Statecraft.Sim/Gdd03ReceptionFixture.cs:23-24@a30013863); its exclusions at :36-54 and the status comment's "Remaining integration needs" list do not authorize routing, settlement or timing defaults.

R-CITY-009: The job default (Q-JOBS-1, owner, 15 October 2026). On each housing step, after the refugees become civilians (R-CITY-004), each city whose default is on sets its employed to min(job slots, civilians + craftsmen), in ascending city id, at no Attention cost. Every city's default is on at the start; an accepted AssignWorkers order fixes that city's number and turns its default off for the rest of the game. Each change is a city cause on the Why screen ("Town: N groups in work"). A city's record writes jobs=manual into the state hash only when its default is off.

R-CITY-010: A city the enemy holds pauses (decision D6, owner, 1 November 2026, "Approve all six (Recommended)", answers; #1711). Under gdd-0.3 only, a city is held in a turn when an enemy military unit (a unit with a side other than the player's; a halutz has no side and holds nothing, R-UNIT-001) stands on its settlement hex when steps 5 and 6 run, after step 3: no later step moves an enemy unit, so that is where it stands at the turn's end. An enemy formation that arrives on the hex at step 9 (R-UNIT-009) holds the city from the next turn. That turn a held city pays no R-CITY-002 job income and its R-PROJ-003 projects gain no progress: their records, turns left and carried fraction stay as they were. Its staffed arms workshops pause the same way: no charge and no progress, the cycle kept (R-ARMS-006). The housing step still absorbs refugees and fills default jobs; its people and every id stay, nobody is evacuated, and an objective already awarded stays awarded. A turn in which no enemy unit is on the hex pays and builds as before. A city the map marks lost (#2955) with no enemy unit on its hex still pays and builds: the lost mark is map-only (owner, 8 November 2026: "Earn when empty (Recommended)"). A held city pauses no Food: only an enemy military unit on a field's own hex stops that field's Food (owner, 9 November 2026: "Yes, enemy on field only (Recommended)", #1711). The ledger harvest (Pools.Harvest given a month, R-FIELD-002, #2972) links no field to a city: a field adds nothing when an enemy military unit stands on its hex, the test Held uses for cities (Cities.EnemyHexes), and an enemy unit on a city's settlement hex stops no field's Food. The gdd-0.3 field fixture pauses a field the same way (R-FIELD-002, #2471): a field not Israeli-held at settlement adds no Food, and its farmer and assignment stay until it is retaken. gdd-0.1 and legacy are unchanged, and gdd-0.1 keeps its R-END-003 defeat. Each turn a city is held, step 6 writes one Why line for it ("Town: held by the enemy; no job income or project work this turn", no change of its own), in city id order; the turn-report panel lists those lines in a row of their own, "Cities held by the enemy", after the stocks and the arms workshops and before the enemy's orders (#1711). A retaken city gets no line.

R-CITY-005: The cities' employed share (employed over civilians and craftsmen, summed over all cities; 1 with none) is an effect-web source of morale, Social cohesion (work_lifts_cohesion, source id employed_share), read after the turn's housing. The legacy loop has no cities, so there it reads 0 and adds nothing. Housing reaches Social cohesion through the existing housing link (homes_lift_morale) until Phase 3 makes it the cities' aggregate (Q-NAT-2, decided 15 October 2026). Phase 1's separate cohesion value and its two weights are gone (D-C1, Reading 2).

R-CITY-006: The cities are the city and town rows of map/settlements_1948.yaml, in that file's order, and their start values are cities.yaml's. The loader refuses a cities.yaml whose ids are not exactly those rows in that order (#2732). Each of a city's values (people, employed_share, housing_capacity, job_slots) is required and has a status: sourced (read from the source named), derived (computed from sourced values), wikipedia-only (an owner-accepted Wikipedia figure) or design (a labelled design value, R-LOOP-005). A sourced, derived or wikipedia-only value has at least one source, each with a url and a checked date, and the day (YYYY-MM-DD) or month (YYYY-MM) it describes; the loader refuses a value without them, naming the file and field (GDD Phase 3 plan, review focus 3).

Both GDD identities read optional regions.yaml (schema/regions.schema.json), separate from aliyah origins, as immutable lookup data (R-ENEMY-007); the regional foundation's population and control state is exact gdd-0.3 only. Each explicit region has a stable id, one label place (capital_location), land member hexes and member place ids. A member place is a settlement, a named map location, a designated site (sites.yaml) or a locality, whose hex is map/locality_hexes.yaml's (written by tools/locality_hexes.py from map/localities_1948.yaml and map/region_places_1948.yaml; read only with regions.yaml, under both GDD identities, so it is part of their content hash; the state hash carries regional state only under exact gdd-0.3). Owner, 2 November 2026: Q1 "Accept localities", and "Fix all, then ship (Recommended)", whose option read "Accept site ids too; author records for the 4 (sourced); for the 14, the member's hex joins its region where it's unclaimed, and where another region holds the hex the member moves to that region. All listed for Amir to review." The label place must be a member; it is a label only (Q3, owner, 2 November 2026: "All nine drafts as labels"), not the regional-capital model the owner declined on 30 October 2026 (GDD, "Regional-capital model"), and nothing gives it capital mechanics. All nine shipped labels are APPROVED as labels (owner, 2 November 2026: "Approve all nine as labels (Recommended)"); that changes no rule. A region may list optional field_hexes, the opening field hexes (#1691 Q4, owner, 2 November 2026: "Approve as drafted (Recommended)", the 20 by the 27 October 2026 split): each must be a land hex of the region's own member hexes, listed once; a hex is in at most one region, so no field hex is in two. They are data only: no rule reads them yet, and the four checks of the shipped list (land, a coastal plain, lowland or valley biome, Israeli-held on 15 May 1948 in map/control_hexes_1948_49.yaml, in its region) are in tools/tests/test_region_field_hexes.py; a person's check of each row is pending. Validation uses effective placements after overrides; a place id with different hexes in two place files, repeated members and overlapping ownership are errors. Different place ids may share one hex within a region. No contiguity or complete-map coverage is required: a land hex may be in no region. Unknown opening control stays null and never grants access. Geography approval is separate from evidence status and human historical review. The shipped file has the nine regions of #1691; no new game initializes them yet (#1689).

The approved region associations and capital choices are recorded in research_notes/gdd03-regions-capitals-fields.md and its evidence ledger. The shipped extents are the sub-district design draft (owner, 30 October 2026: "C: sub-district, then Amir adjusts") with the 2 November 2026 fixes above, listed for Amir to review; they are design regions, not settled historical extents, and regional activation stays off (#1689). The inactive schema input research_notes/gdd03-regions.schema-input.yaml stays as evidence and retains every member id; the shared strict loader refuses seven of its nine rows, each on a member place that the 2 November fixes resolved. Named-place acceptance projections do not stand in for complete regions. Reception-camp design association does not establish its footprint, operation or control; historical rows still require a person's check.

Regional state is currently initialized only by explicit internal fixtures, through the existing population ledger. Immutable pools keep base ordinary housing, condition fractions, Industry/workshop slots and uniquely owned facility contributions separate; reception/training beds do not add ordinary housing. Population and assignment summaries come from the ledger. Canonical geography/capacity sections are limited to initialized gdd-0.3 regional fixtures, and clones share only immutable records. No shipped era selects regional initialization, no nonempty production regional roster or allocations are supplied, and the aggregate-city guards remain. The city behavior above remains the production path; capital-addressed commands, replay, capacity reservation policy and the production cutover are later increments of #1690 and its consumers.

R-CITY-007: A city starts with civilians = people / people_per_unit to the nearest unit (a half rounds up), and at least 1 when people is above 0 (Q-START-2, owner, 15 October 2026); employed = people * employed_share / people_per_unit to the nearest unit; housing_capacity and job_slots as written, in units. The loader refuses a city whose start employed is above its job slots. The arithmetic is decimal, so it is exact on every runtime. It holds unchanged for the Historical start under gdd-0.3 (D2, owner, 1 November 2026: "Approve all six (Recommended)"; docs/GDD.md:348): the census cities count at 1,000 people a unit, each city's employed is its people times its share, and the starting ledger holds one census record per start unit (HistoricalPopulationTests.RealInventoryHoldsTheCensusCitiesAndTheStartingFormationsNoMoreNoLess, HistoricalPopulationTests.CensusSourcesAreTheTwelveCitiesAtTheirCensusUnits, HistoricalPopulationTests.TheHistoricalStartCountsAtOneThousandPeopleAUnit, HistoricalPopulationTests.EachCitysEmployedIsItsPeopleTimesItsShare).

R-HOUSE-001: Each city has a housing condition from 0 to 1, an average over the city (D-D2, Q-HOUSE-2, owner, 15 October 2026). Every city starts at mvp.housing_start_condition (Q-START-3: one design value for all the cities, 0.35, model.yaml's housing base). The condition is per city: one city's counts never change another's. It is in the state hash (R-LOOP-002).

For R-HOUSE-001, the preceding legacy/gdd-0.1 wording remains unchanged. The inactive gdd-0.3 cross-reference is Gdd03HousingFormula, detached synthetic arithmetic with an explicit condition input, no initial condition or state commit (sim/Statecraft.Sim/Gdd03HousingFormula.cs:5-6@7ec122145). It snapshots labelled Agreed/Default catalogue scalars, preserves the Default basis of overflow deterioration, and refuses missing, Deferred, unlabelled or malformed required scalars and incompatible catalogue identity. Condition is bounded to 0..1 after deterioration and each explicit project effect; repair uses the catalogue's condition-pp part without implementing repair cost, work or availability (sim/Statecraft.Sim/Gdd03HousingFormula.cs:13-62@7ec122145, :80-103; sim/Statecraft.Sim.Tests/Gdd03HousingFormulaTests.cs:78-98@7ec122145, :119-141). Source: docs/gdd/cities.md:27@605277dfa511dc512562ececeb273b8393b21d35, "Project capacity/condition effects live in [balance](balance.md#project-catalogue), clamped to condition bounds." (quoted verbatim from that pinned file; its balance.md is not in this repository) Initial-condition integration remains pending: the #1703 status comment of 27 October 2026 leaves initial-condition and resident-policy questions with #1668, "rather than becoming fallback values". Under gdd-0.3 the live game now starts every city at the same mvp.housing_start_condition 0.35 (27M-16, owner-approved 1 November 2026: "Uniform: one value for every city, 0.35."), and the condition moves by R-HOUSE-005 (#1703).

R-HOUSE-002: A city's overflow is max(0, housed - housing_capacity), one capacity point housing one population unit (D-D1, Q-HOUSE-1, owner, 15 October 2026). On the housing step, after R-CITY-004 and before Social cohesion (D-A6), a city with overflow loses overflow_condition_loss * overflow / housed of its condition, never below 0; a city without overflow keeps its condition. overflow_condition_loss is 0.15, a design value. The loader requires both values, from 0 to 1. Overflow reaches Social cohesion through R-HOUSE-004 (a lower condition lowers the base of the national housing indicator) and the existing homes_lift_morale link (k 0.15) from housing into morale, which is Social cohesion (R-NAT-001); there is no direct overflow term (Q-HOUSE-1, decided by the owner 15 October 2026, plan:165 of docs/superpowers/plans/2026-10-15-gdd-phase-3.md: cohesion follows through the existing housing to morale link, "so it needs no number of its own"; the cohesion half was built after #728, merged as #857, owner 17 October 2026).

For R-HOUSE-002, inactive gdd-0.3 synthetic arithmetic prices the surcharge as SurchargeResidents * max(0, money_threshold - ConditionBefore) * money_slope, then subtracts the overflow loss and clamps condition. Overflow loss is (overflow_deterioration / 100) * max(0, Residents - Capacity) / Residents, with Residents the housed residents after the step's arrivals become civilians (27M-15, owner-approved 1 November 2026: "After that step's arrivals become civilians (R-CITY-004)."; #2517); zero residents gives zero loss. There is no unpaid-upkeep loss and the helper takes no upkeep input: there is no housing upkeep (27M-17, R-HOUSE-005), and housing.max_deterioration stays in content/gdd03-balance.yaml as post-MVP design that nothing reads (#2516). The returned cohesion contribution is max_cohesion_loss * (CohesionResidents / OwnPopulation) * max(0, (cohesion_threshold - ConditionAfter) / cohesion_threshold), zero with no own population. It is displayed points, not an indicator write or a national summation (sim/Statecraft.Sim/Gdd03HousingFormula.cs:66-88@25938871c; the after-arrivals, worked-number, distinct-count, post-deterioration and zero-denominator fixtures in sim/Statecraft.Sim.Tests/Gdd03HousingFormulaTests.cs:77-86,112-172@25938871c). Source: docs/gdd/cities.md:29-35@605277dfa511dc512562ececeb273b8393b21d35; #1703 says, "Price the surcharge from pre-deterioration condition, then deteriorate once, then apply cohesion once." Under gdd-0.3 this arithmetic is now live, per city, as R-HOUSE-005 (#1703); under gdd-0.1 and legacy the preceding active rule (0.15) is unchanged. Repairs are R-HOUSE-006 under gdd-0.3 (#1700); monthly snapshot timing (#1716) remains pending.

R-HOUSE-003: The people who take a city's housing (housed) are its civilians, its craftsmen, its recruits and its groups in training, a halutz group included until it becomes a movable unit. Halutz units do not (Q-HOUSE-5, owner, 15 October 2026). Refugees wait for housing and are not in it. With an initialized Historical ledger (the dark start, R-POP-002; #1725 PR 7) the housed people are counted from the ledger's records, one population unit each, not the aggregate city fields: a city houses its free civilian and trainee records located in it, so the opening field farmers are housed in their host city beside the place each adds (R-POP-004). Nir Am and Nirim have no city record, so their two farmers are housed in no city.

For R-HOUSE-003, the inactive gdd-0.3 helper takes independently supplied surcharge, overflow and cohesion resident counts plus own population, the overflow count taken after the step's arrivals (27M-15); it selects no residents and substitutes none of those counts for another (sim/Statecraft.Sim/Gdd03HousingFormula.cs:66-88@25938871c, :130-150; sim/Statecraft.Sim.Tests/Gdd03HousingFormulaTests.cs:77-86,135-144@25938871c). Negative counts or capacity, cohesion residents above total own population, and condition outside 0..1 are refused. A zero-population cohesion input therefore requires zero cohesion residents (docs/gdd/cities.md:27,31@605277dfa511dc512562ececeb273b8393b21d35). Source: docs/gdd/cities.md:25@605277dfa511dc512562ececeb273b8393b21d35, "A civilian resident without dedicated reception, teacher or military-training housing uses ordinary housing."; docs/gdd/loop-switch.md:33 at that revision said, "use pre-arrival residents for overcrowding deterioration", which 27M-15 superseded: housing effects count residents after the step's arrivals become civilians (owner-approved 1 November 2026; owner, 2 November 2026: "After housing arrivals (Recommended)"), so Gdd03SyntheticMonthContext no longer carries a pre-arrival resident list (#2517). Those are producer requirements, not resident-selection behavior implemented here. Live status selection (#1689), regions/approved housing locations (#1691; #1690 closed on 2 November 2026, its post-MVP regional pools moved to #2233) and which residents to count (#1668) remain pending under the #1703 status comment, which says, "Regional housing settlement is not yet implemented". The preceding legacy/gdd-0.1 resident set remains unchanged. Under gdd-0.3 (#1703) a city's Israeli formations are not housed: housed is its civilians, craftsmen, recruits and groups in training, so the formations whose home it is no longer count toward its overflow, R-HOUSE-004's weights or R-OBJ-004 (27M-18, owner-approved 1 November 2026: "Housed = civilians, craftsmen, recruits and groups in training. Refugees and pioneer units are not housed."; confirmed by the owner, 3 November 2026: "Yes, formations not housed (Recommended)"). gdd-0.1 keeps counting them (Q-HOUSE-5; Cities.Housed).

R-HOUSE-004: In the GDD loop the base of the national housing indicator is the people-weighted mean of the cities' housing condition, each city weighed by its housed people (R-HOUSE-003): sum(condition * housed) / sum(housed). Policy effects, map actions and events still act on top, as on any indicator. Decided (owner, 17 Oct 2026, option (a) of Q-PR3-A, relayed on #855 at 06:33Z): the national housing base is the people-weighted mean of city condition, with policy, map actions and events still applied on top. When no city houses anybody the mean is undefined, and the base is model.yaml's. Capacity reaches the level through overflow and condition (R-HOUSE-002).

R-HOUSE-005: Under gdd-0.3 only (#1703, the MVP's per-city housing condition), the housing step after R-CITY-004 settles, in #1703's order ("Price the surcharge from pre-deterioration condition, then deteriorate once, then apply cohesion once."), for every city with a condition, with R its housed people (R-HOUSE-003) and P all own population (every city's housed people and refugees plus the strength of every player unit, halutz units included; with an initialized Historical ledger, every city's housed and refugee records plus the active soldier records, so a record in no tracked city, such as a farmer at Nir Am or Nirim, is not in P, as the owner's 2 November 2026 answer "The 12 cities' people (Recommended)" keeps P to "only the cities whose housing is tracked"; #1725 PR 7): (1) a surcharge of R * max(0, money_threshold - c) * money_slope Money a turn from the condition c before this turn's deterioration, summed over the cities into one Money line (housing_surcharge) and borrowed when the Treasury is short (R-MONEY-003); money_threshold is 0.50 and money_slope is 0.1 Money per housed population unit (owner, 2 November 2026: "Scale it to unit size (Recommended)"; the option read: "0.1 per unit (as #2112 did for Industry), about 7.6 Money a turn at start."), so the shipped census start charges 511 housed units x 0.15 x 0.1 = 7.665 Money on the opening turn; (2) overflow deterioration of overflow_deterioration * max(0, R - housing_capacity) / R percentage points (5 pp; owner, 2 November 2026: "Yes, 5 pp under GDD 0.3 (Recommended)"), never below 0, in place of R-HOUSE-002's overflow_condition_loss; zero housed loses nothing; (3) a Social cohesion loss of max_cohesion_loss * sum((R / P) * max(0, (cohesion_threshold - c') / cohesion_threshold)) displayed points (max_cohesion_loss 2, cohesion_threshold 0.25; owner, 2 November 2026: "Keep the 2-point loss (Recommended)") from the condition c' after deterioration, applied once to morale through its permanent shift with one cause line (housing_cohesion); zero own population loses nothing. The legacy homes_lift_morale link from housing into morale adds nothing under gdd-0.3, so housing lowers Social cohesion only through (3) (#1703 Done-when, with #1715; the Food term stays #1715's). There is no housing upkeep (27M-17, IL 0m), so no unpaid-upkeep deterioration. Values: content/gdd03-balance.yaml (housing.*), read by Gdd03HousingFormula; code: Cities.UpdateCondition and Cities.SettleGdd03, EffectsSystem.RetiredGdd03Link; tests: Gdd03HousingConditionTests, Gdd03MoneyLedgerTests. The issue's three rules run under gdd-0.3: the 5 pp overflow (owner, 2 November 2026: "Yes, 5 pp under GDD 0.3 (Recommended)") and the cohesion term (owner, 2 November 2026: "Keep the 2-point loss (Recommended)"; the option read: "As issue #1703 specifies, GDD 0.3 only."). P is the 12 cities' people as built, plus the player units (owner, 2 November 2026: "The 12 cities' people (Recommended)"; the option read: "As built: only the cities whose housing is tracked."). R is counted after the step's housing arrivals (owner, 2 November 2026: "After housing arrivals (Recommended)"; 27M-15). Formations are not housed (owner, 3 November 2026: "Yes, formations not housed (Recommended)", #1703); repairs are R-HOUSE-006. On an initialized Historical ledger (the dark start, #1725 PR 7) the step runs as above, with R and P counted from the ledger's records (R-HOUSE-003), and changes no record; a synthetic or Open ledger is still refused (R-POP-002). The halutz-unit case is not handled yet on a ledger game (R-HOUSE-003, R-HOUSE-005): no record is a halutz unit at the dark start, and the ledger counts have no halutz rule, so a later halutz-unit record would be counted only by its status and city, not as a halutz unit. The city panel shows each city's terms on the state as it is now (Cities.HousingOutlook, #1722): the same formula, with that city's R and the current P and condition, changing nothing; the turn's own step can differ, as the step's arrivals and finished repairs come first.

R-HOUSE-006: Under gdd-0.3 only (#1700, #1703), a housing repair is the city project housing_repair (content/projects.yaml): "A repair costs 0.5 Money and 1 work and raises the city's housing condition by 10 percentage points" (docs/GDD.md, "Housing repair price"; owner, 1 November 2026: "Fixed price 0.5 Money + 1 work"; labelled design values). Its row matches content/gdd03-balance.yaml housing_repair, which the loader checks: cost 0.5, work 1 as turns: 1 (maintainer call (not an owner ruling): one turn of a builder's work, as arms_workshop's work 3 is its 3 turns), condition_pp 10 as condition: 0.1, and no capacity or job slots. It is started, paid, built and cancelled as any city project (R-PROJ-001 to R-PROJ-003, R-CITY-010, R-CITY-011): the Treasury pays 0.5 at the start and borrows what it lacks (R-MONEY-003), it uses the turn's city start (one per city, plus map_actions_per_turn 2 nationally), it needs a builder to gain its turn of work, the housing_construction multiplier (the Labour minister's, content/rules.yaml) scales that work as for every housing build (R-MIN-002; maintainer call (not an owner ruling): a repair counts as a Housing build, so with a weak minister it can take more than one turn), and its builder is freed when it is done or cancelled. When done it adds 10 percentage points to its city's condition, never above 1. A repair is done on the city itself and has no hex of its own (owner, 3 November 2026: "No hex needed (Recommended)"; the option read: "Repairs are done on the city itself; they fix existing housing." #1700). This departs from Q-PROJ-2 for repairs only: a repair is ordered on its city's settlement hex and nowhere else (refused elsewhere as "A housing repair is done on Town A's own hex"), whatever placement or city project that hex holds, and takes the hex from no other project; once done its record leaves the state, so it leaves no building and the hex still shows its own project. Every other kind keeps Q-PROJ-2's free hex. The repair lands on step 5, before R-HOUSE-005's deterioration on the housing step, so a city with overflow can gain less than 10 points in that turn. gdd-0.1 and legacy have no repair: the loader keeps the row out of them.

City projects

A city project is built in a city (GDD "Cities and construction"; Phase 3 PR 4, #735). The kinds are in content/projects.yaml, every number a design value (Q-PROJ-1, Q-PROJ-5, owner, 15 October 2026, #375: "Take all recommendations", with agent-proposed values the owner may override): permanent housing (housing), the transit camp (transit_camp, D-D3, Q-HOUSE-3) and industry, and, under gdd-0.3 only, the arms workshop (arms_workshop, R-ARMS-004) and the housing repair (housing_repair, R-HOUSE-006). They replace the legacy map actions, which the GDD loop refuses (R-PROJ-004).

R-PROJ-001: StartProject(kind, city, q, r) starts a city project of a projects.yaml kind for a city on hex (q, r). The Treasury pays the kind's cost at once and borrows what it lacks (Q-PROJ-1, Q-MONEY-3, "Allow borrowing"); the project takes the next id and the kind's turns, the turn of the order the first (D-A3). It needs no workers (Q-PROJ-3); under gdd-0.3 the start takes a free person as its builder when there is one (R-CITY-011). It is refused, state and Attention unchanged, for an unknown kind or city; a locked kind that no event has unlocked (R-EFX-008); a transit camp without the maabarot policy at 0.5 or more (Q-HOUSE-3: the policy carries the camp's sourced date, and the map actions' 1 January 1950 window does not apply); a hex that is not land, is more than map_action_max_distance from the city's settlement hex, or holds a map-action placement or another city project, built or not (Q-PROJ-2; a gdd-0.3 housing repair is on its city's own hex and needs no free hex, R-HOUSE-006); and once map_actions_per_turn projects were started this turn, counted nationally (Q-PROJ-4; the GDD loop builds no map action, R-PROJ-004). The legacy loop refuses it before charging Attention. Its Attention cost is attention_cost.start_project (1, Q-PROJ-5). Under gdd-0.3 (owner, 3 November 2026: "One per city, plus 2 nationally") a city starts at most one city project a turn: a second StartProject for the same city that turn is refused with "Town A has started a city project this turn", with Money, debt, builders and every project unchanged. The national map_actions_per_turn count still applies. The city's refusal (city_cap) is checked just before the national one (turn_cap), so a start over both gives the city's reason. The count is kept per city for the turn of the national counter (GameState.CityStartsThisTurn), is hashed as cstart{...} only when it has a start, and is never kept under gdd-0.1, so no gdd-0.1 hash changes (#1700 slice B).

R-PROJ-002: CancelProject(id) removes a city project that is being built. The Money paid is not returned, and there are no people to return (D-A3, Q-PROJ-3); under gdd-0.3 its builder, if any, is free again at once (R-CITY-011). An unknown or built project is refused with the state unchanged. Its Attention cost is attention_cost.cancel_project (0.5, Q-PROJ-5).

R-PROJ-003: On step 5 of the turn (D-A6), after the pioneer work, every city project being built has one turn of work less to go, in ascending id - 1 x m of one at the system's office multiplier since the ministers (R-MIN-002), the fraction carried on the record as on R-SITE-003. One that reaches 0 is built on the turn's end date and adds its kind's capacity to the city's housing capacity, its condition to the city's housing condition (kept from 0 to 1, R-HOUSE-001) and its job_slots to the city's job slots, once: a built project keeps its record, with its hex, and is never advanced again (a gdd-0.3 housing repair has no hex, and its record leaves the state once done, R-HOUSE-006). GameState.CityProjects is hashed as cproj[...] only when it has a project (R-LOOP-002). Under gdd-0.3, at step 5 a project gains the turn's progress only when it has a builder and its city still has that person without a job (each project with a builder takes one, in ascending id). A project with no builder waits: its record, turns left and carried fraction stay as they were, as in R-CITY-010, and the Why screen says why ("Town: a housing project waits: no builder", or "... no person without a job is left for its builder"). Money is never an input to progress. A project's builder is written into its hashed record (;builder) only when it has one, so no gdd-0.1 hash changes.

R-PROJ-004: In the GDD loop a map action (map_action, the legacy transit camp or housing placement) is refused with the reason "Map actions are not part of this game: start a city project instead" and its Attention refunded, whatever its action or hex (plan PR 4, spec §1.4: "Placements become completed projects"). So no placement is built and none counts against R-PROJ-001's turn limit. The info panel shows no legacy Build button in the GDD loop. The legacy loop keeps the map actions until that code is deleted (D-E1). The pioneer orders from the map (pioneer_work, R-SITE-001) are not map actions.

R-PROJ-005: A specialised project is allowed only in a city that has its resource (D-D8: "A winery requires areas for grapes"; Phase 3 PR 6, #737). A projects.yaml kind may name a requires: resource; a cities.yaml city lists its resources:, each with an id and sources (a url and a checked date each), and may say what the sources leave open (to_verify). StartProject of such a kind in a city without the resource is refused like the other R-PROJ-001 refusals, before the hex checks; a city founded in play has no resources. The loader refuses a resource without sources and a kind whose resource no city has. Otherwise the kind is an ordinary city project (R-PROJ-003). The one such kind is the winery, which needs vineyards, in Rishon LeZion with its vineyards flagged to_verify (owner, 20 October 2026, on #737, as GDD v0.2 says; this replaces the 18 October default of Zikhron Ya'akov, which has no winery).

R-CITY-011: Under gdd-0.3 only (owner, 3 November 2026: "Builders in the MVP (Recommended)", "One builder, today's times (Recommended)", "Yes, if someone is free"; #1700 slice A), a city project being built has at most one builder. A builder is one of the city's civilians and craftsmen who hold no job and are not a staffed arms workshop's worker or another project's builder (a free person). The builder stays one of them, with no job, profession or qualification change, is not employed (R-CITY-001) and earns no job income (R-CITY-002). Train, ReassignPioneers, a job count, the job default (R-CITY-009) and a workshop's worker cannot take a builder. StartProject takes the city's free person as the new project's builder when it has one, at the moment the start runs, in queue order: a Train or AssignWorkers queued before the start has already changed the count, and one queued after it sees the builder as taken. Otherwise the project starts with no builder; it never takes anyone out of a job, the job default's included, and a project that waits does not pick up a person who becomes free later. On the aggregate cities the builder is one unit of the city's count, counted at the ordinary tier (no cost in the MVP). AssignWorkers(city, "build", n) sets the number of the city's projects being built that have a builder: a rise puts builders on the projects that have none, in ascending project id; a fall takes them off the projects that have one, in descending project id. It is refused, with the state unchanged, for an unknown city; a negative count; more than the city's projects being built ("Town A has 1 city project being built"); or more new builders than the city's free people ("... civilians and craftsmen without a job"), with R-ARMS-005's job-default rule (while the default is on, an order may take a builder from the default's jobs, and is refused only past the city's civilians and craftsmen less its staffed workshops and builders). Under every other ruleset the role is refused as "There is no such role". A project's builder is freed when it is built or cancelled. A freed builder does not move to another project by itself. The limits a panel reads with this turn's queue (Training.Available, Workshops.Check, Cities.MaxWorkers) also count the builder each queued StartProject in the city may take, so a count up to them is accepted whichever order runs first. On step 6 the staffed arms workshops share the city's people without a job with the builders, who took theirs at step 5 (R-ARMS-006). The Why screen says when a builder starts on a project ("Town: a builder started on a housing project") and when one is freed, as the project is built or cancelled or the build count falls ("Town: a builder left a housing project and is free again", #1700 slice B).

For R-PROJ-001 and R-PROJ-003, the preceding rules and the IL 0m upkeep of R-MONEY-002 remain unchanged in the game. The inactive gdd-0.3 addition is Gdd03ProjectBuildFixture, synthetic builder rules over the #1963 project catalogue (#1700; no command, save, loader, resolver or UI caller). A start pays the kind's full catalogue cost up front; when Money is short the start goes ahead and the shortfall is borrowed as R-MONEY-003 says, Money stopping at 0 and the rest moving to the fixture's debt (owner, 1 November 2026, D1: "Approve all six (Recommended)"). A project has at most one builder, and each turn with its builder counts as one turn of the kind's turns in content/projects.yaml (owner, 3 November 2026, #1700: "One builder, today's times (Recommended)"); a project with no builder makes no progress, and an order naming one project twice is refused with nothing changed. The record holds the turns its builder has worked. A catalogue kind with no row of the same id in content/projects.yaml (today all but industry and arms_workshop) has no turns to count, so its start is refused before anything is paid (maintainer reading). Money is not an input to progress, so extra Money never changes the completion turn (Amir, 28 October 2026, recorded on #1668: "Extra Money cannot replace work or speed completion. Housing and improvements also require Money for upkeep."). The project completes on the turn its builder works the kind's last turn; the builder is then released, and an order to a completed project is refused. From the completion turn on, every turn, a completed project is charged R-MONEY-002's MVP rate, IL 0m, vacant or not, so Money does not change; the catalogue's non-zero upkeep is post-MVP and unused (owner's 30 October 2026 "Keep ruling" on UPKEEP, kept on #1700 on 1 November 2026). A city starts at most one project a turn, and two starts are allowed a turn nationally, as R-PROJ-001 under gdd-0.3 (owner, 3 November 2026: "One per city, plus 2 nationally"): a second start in the same city, or a third in the country, is refused before anything is paid, the city's limit first, with Money, debt and every project unchanged, and both counts reset when the turn ends; projects already running keep progressing. The catalogue's construction.starts_per_turn (1, "nationwide per turn") is Amir's catalogue reading; the live rule is R-PROJ-001's. The fixture has no pioneer start, so pioneer work (#1713) never uses the allowance. Cancel, housing repair, completion effects and the live integration are other #1700 slices.

Then changed (3 Nov 2026, #1700): the owner chose the option labelled "One builder, today's times (Recommended)". A project has at most one builder, and a turn with its builder counts as one turn of the kind's existing turns, as R-CITY-011 and the gdd-0.3 parts of R-PROJ-001 to R-PROJ-003 build it in the game; the builder tiers, their rates (1, 1.5, 3) and builder.slots_per_project (3) above are set aside for the MVP. Gdd03ProjectBuildFixture follows this since #2514.

Training

Training is step 4 of the turn in the approved order (§2.3, D-A6): it runs before production. One military camp and one Hakhshara exist in Tel Aviv at the start, a labelled design assumption (D-D5); pioneers build more at camp sites (R-CAMP-001). Scope note: this is the live gdd-0.1 start. Under gdd-0.3, 27M-25 says: "Schneller and Tel HaShomer operate from the start, the other military camps are build sites, every camp uses military.camp_capacity 5, and the reception site is Kiryat Shmuel, Haifa" (that 5 is #1668's value at 10,000 people a unit; at D2's 1,000 people a unit it is 50 units, content/gdd03-balance.yaml:307). Only the dark start enforces that camp capacity (R-TRAIN-001): a gdd-0.3 game started without one (the public Simulation.NewGame) still runs the aggregate Train, which does not read military.camp_capacity, the gdd-0.1 behaviour until #1689's last slice (#2528). R-CAMP-003 defines the gdd-0.3 start camps. The Hakhshara and its Train track stay under gdd-0.3 (owner, 4 November 2026: "Keep the Hakhshara", #2484).

R-TRAIN-001: Train(city, track, n) takes n of the city's civilians without a job (not a staffed arms workshop's worker, R-ARMS-005, or a builder, R-CITY-011) into a cohort that trains as track (a unit type with a facility: recruit, artillery and armoured_car in a military camp (R-UNIT-008), halutz in a Hakhshara, craftsman at industry, R-CRAFT-001) for the type's training_turns, counting the turn it was ordered in. The Treasury pays n * training_cost at once. When the Treasury holds less, the shortfall is borrowed, as for map actions: it is added to the debt and the Treasury stops at 0 (owner decision "Allow borrowing", 15 October 2026). The trainees leave the civilians at once; the cohort takes the next id, which units share. It is refused, with the state unchanged and the Attention refunded, when the city or track is unknown, n is below 1, the city lacks the facility, or it has fewer civilians without a job. A city has a facility when cities.yaml gives it one at the start, gdd03-start.yaml gives it one under R-CAMP-003, or a camp site of that kind built for it is in the state (R-CAMP-001). The legacy loop refuses it before charging Attention. Reinforce (R-TRAIN-004) starts a cohort by the same rules, aimed at a unit. A queued ReassignCraftsmen or ReassignPioneers (R-QUAL-004, R-QUAL-005) takes its civilians without a job as a queued Train does (counted in full, so the orders are accepted in any order), in Training.Ordered and Training.Available (#1548; owner, 24 October 2026: "Yes, count them", https://github.com/eyalgolan/statecraft-israel/issues/1511#issuecomment-6378360813). Under gdd-0.3 the halutz track still trains in a Hakhshara (owner, 4 November 2026: "Keep the Hakhshara", #2484). It runs on a game with an initialized Historical ledger (the dark start, R-POP-002; #1725 PR 7, producer 12) under the owner's rulings of 8 November 2026 (#1725 issuecomment-6724941867). "Recruit+crews+halutz (Rec.)": the recruit, crew and halutz tracks run, and the craftsman track is refused with a plain reason until #2476, the reason "Craft training does not run on the population ledger yet (#2476)". "Cohorts stay main (Rec.)": State.Cohorts stays the record of truth, each trainee's Course target is per facility (military_camp/<city>, hakhshara/<city>), and the cohort keeps a sorted list of its members' ids, hashed only when set. "Unskilled first, keep (Rec.)": civilians with no profession first, then the qualified, each in ascending id, and a qualified trainee's profession is kept dormant, not cleared. "Keep costs, pool rest (Rec.)": one item is issued per record, and the rest of training_arms is taken from the stock and kept in Cohort.Arms, so a cancel returns it. "Enforce it here (Rec.)": the camp capacity, 50 records per camp city over all military tracks, is enforced here, refused with "'s military camp has room for more trainees (capacity 50)". "Start + runtime bind (Rec.)": the start facilities' Course targets are bound at the start, plus one all-or-nothing runtime bind for new formations and built camps. There the trainees are the city's free-civilian records without an assignment, located in the city; a refugee, a worker and a farmer on a field are not idle. Each becomes a trainee on its facility's course in the same place (27M-7), and each military trainee is issued one military item, training/<cohort>/<id>, from the Arms stock. Every check (city, track, count, facility, idle records, camp capacity, Arms) comes before the first write, and the records and any Course target a built camp still needs are validated as one candidate. The Money, the cohort id and the Why lines are the aggregate game's; no aggregate city field is written. The ledger also binds each military track's unit class as a qualification at Initialize. A trainee who already carries a military item, and a farmer or craftsman on the halutz track, are refused before any write, the state unchanged (owner, 8 November 2026, Q3: "Refuse, with reason (Recommended)"), with "A civilian without work in already carries a military item" and "A farmer or craftsman in cannot train as a halutz"; neither is idle on the dark start, and a replacement cohort (R-TRAIN-004) is refused the same way. These ledger refusals are internal English text: their Hebrew comes with the UI wiring PR (owner, 8 November 2026, Q4: "With the UI PR (Recommended)"). A synthetic or Open ledger is still refused.

R-TRAIN-002: CancelTraining(cohort) returns the cohort's people to its city as civilians without a job and refunds none of the Money paid (D-A3); a recruit cohort's Arms go back to the pool (R-ARMS-003). A cohort that does not exist, or has finished, is refused with the state unchanged. Replacements (R-TRAIN-004) cancel the same way. On a Historical ledger game (R-TRAIN-001) the cohort's member records are free civilians without a job again, in place, with their reception, and only the items this course issued leave them; those items and the cohort's pooled Arms go back to the stock, as many as the aggregate cohort's Arms.

R-TRAIN-003: Each turn, in ascending id, every cohort has one turn less to go. A cohort that reaches 0 ends: a recruit cohort becomes one military unit with its count as strength (R-UNIT-003), and so does a gun or armoured car cohort, of its own type (R-UNIT-008), craftsmen join the city's count (R-CRAFT-001), and halutzim become one movable unit per group. Units appear at the city's settlement hex, taking the next ids in turn. New recruit completions never fill City.Recruits; its old canonical field is retained, and saves still replay their command log. Cohorts are written into the state hash only when there are some. A cohort aimed at a unit ends as R-TRAIN-004 says instead. Reassignment of a qualified civilian is a second way into the role, with no cohort (R-QUAL-004, R-QUAL-005). On a Historical ledger game (R-TRAIN-001) each turn of a cohort counts one course tick on each of its records, and every record change and new target of the step is validated as one candidate before any write. Owner, 8 November 2026, "New unit at graduation (Rec.)" (#1725 issuecomment-6724941867): a new formation target is bound at graduation, one new formation per cohort (R-UNIT-003). A finished recruit or crew cohort's records are active soldiers serving the new target formation/<unit id> in the city's region, with the type's class as their qualification and their item kept; like an opening soldier each has the region-only location of its formation's region (owner, 8 November 2026, Q2: "Formation region (Recommended)"). "Free pioneer, not housed (Rec.)" (#1725 issuecomment-6724941867): a halutz graduate is a free civilian, Pioneer, Novice and unassigned, in the city's region with no city, so the city no longer houses it; a record that is a pioneer already keeps its tier and work XP (owner, 8 November 2026, Q1: "Keeps tier and XP (Recommended)"). The map units are made as above. On a Historical ledger game each halutz map unit keeps the ids of the pioneer records that form it (owner, 9 November 2026: "Add the link first (Recommended)"): Unit.Records, one per group, the cohort's records in ascending id for its units in id order. "List on the unit (Recommended)": the list is sorted and written into the state hash as |records= only when set, like a cohort's members, so only a ledger game with a trained pioneer unit hashes differently; the records stay free, unassigned pioneers. "Yes, keep the link (Recommended)": a veteran pioneer made by promotion (R-VET-001) keeps the list, as every other change of a unit does.

R-TRAIN-004: Reinforce(unit, city, n) (#1510; owner, 23 October 2026: "Replacements cost Arms", https://github.com/eyalgolan/statecraft-israel/issues/1510#issuecomment-6372356139, and "Default (Recommended)") refills a damaged formation of the player's with trained replacements. The unit must be a military unit of the player's side standing on city's settlement hex. Its course is the trained type of the unit's side and class (a type with a facility): the recruit's basic training for a recruit or a veteran soldier, the guns' and the armoured cars' own for artillery and armoured cars (R-UNIT-008); a type no course trains (any enemy type) takes none. The cohort then follows R-TRAIN-001: the city's facility, n of its civilians without a job, n * training_cost with borrowing (R-MONEY-003), and the Arms of R-ARMS-003 for that course; it carries the unit's id, written into the state hash (|unit=<id>) only when set. The cap is the strength the unit had before its first loss, Unit.Full: combat stores it on a player's unit at its first loss (R-COMBAT-003) and clears it when replacements bring the unit back to it, and it is written into the state hash (|full=<n>) only when set, so a unit that has lost nobody, and every enemy unit, hashes as before. The order is refused, the state and Attention unchanged, for an unknown unit ("There is no such unit"), a unit with no side or an enemy unit (R-ORDER-001's wording), an unknown city, a unit off the city's hex ("Unit N is not in "), a type no course trains ("Unit N cannot take replacements"), a unit with nothing to refill ("Unit N has no losses to replace") and n past the cap less the unit's strength and the replacements already training for it ("Unit N can take K more groups"), then for R-TRAIN-001's and R-ARMS-003's reasons. A queued Reinforce takes its city's civilians as a queued Train does (R-CITY-001). In step 4, before a turn is counted, replacements whose unit has left the state (destroyed, captured, sent home) or stands off the city's hex are cancelled as R-TRAIN-002 cancels a cohort: people and Arms back, the Money spent, with a Why line. A cohort that finishes adds its count to the unit's strength and makes no unit and takes no id; the unit keeps its type, service, order and home (a veteran stays a veteran, R-UNIT-004). The people join the unit's home city's housing (R-HOUSE-003). attention_cost.reinforce is 1, Train's cost, a design value (owner, 24 October 2026: "1, like Train (Recommended)", https://github.com/eyalgolan/statecraft-israel/issues/1510#issuecomment-6375591436). The save line is reinforce <unit> <city> <n> and the scenario form { kind: reinforce, unit, city, count }. On a Historical ledger game (R-TRAIN-001) the owner's ruling of 8 November 2026, "Full minus strength (Rec.)" (#1725 issuecomment-6724941867): the room is Unit.Full minus strength minus training, and a formation whose strength differs from its serving records is refused, the refusal "Unit has soldiers on record but strength ". A starting formation's records serve units_1948/<row> (unit id i is c.StartingForces[i]), a trained one's formation/<unit id>. Finished replacements serve the unit's own target, as active soldiers with the course's class as their qualification.

R-TRAIN-005: In gdd-0.3 synthetic fixtures only (#1697; nothing in the game calls it), a farmer or pioneer course starts at an operating training centre, a facility in its host city with no reception camp, centre upgrade or reception bed (owner, 1 November 2026, "No reception-bed layer (Recommended)"; docs/gdd/training.md:38, "The MVP has no reception-bed layer"). A centre has one veteran teacher and one pupil, the teacher housed in the centre's city (GDD, 1 November 2026 answers, docs/GDD.md:372: "One veteran teacher and one pupil per centre, the teacher housed in the centre's city (Q42)."). A population location has a region and an optional hex, facility or city (R-POP-003; the Historical fixture names each census record's city, R-POP-004), but this fixture does not read the city yet, so the centre's region stands in for its host city, an approximation until it does (#1691). A teacher outside the centre's region is refused with a plain reason that cites Q42. Q42 rules only on the teacher; that the trainee lives in the centre's host city too was our reading (#1697 body: "the teacher and the trainee are housed there as ordinary residents"; docs/GDD.md:374, "Training is ordered in the camp's city, with no travel (27M-7)."), and is now the owner's ruling (owner, 4 November 2026, Q23: asked "Must a farmer-course or pioneer-course pupil live in the training centre's host city?", the owner answered "yes", recorded on #1697). The fixture still uses the centre's region as the stand-in for its host city, so a trainee outside the region is refused with a plain reason that says so; the rule tightens to the host city once centres are bound to real cities (#1691). A centre takes at most one teacher and one pupil, whatever profession.trainees_per_teacher says: a second teacher, or a second pupil while one is on a course there, is refused with a plain reason that cites Q42, not as Deferred, and an inventory with either is refused the same way. The teacher is a free civilian, a veteran of the matching profession; the trainee is a free civilian who is a novice of that profession, or a free civilian with no profession (#2464; docs/gdd/training.md:48: "An eligible free civilian may train under the retained course rules."). That such a pupil stays without a profession on the course is a maintainer reading, not an owner ruling: #1697, update of 3 November 2026, "the MVP has no novice stage for farmers or pioneers: an entrant is a free civilian with no profession label until completion, and a cancel leaves an ordinary civilian", built on the owner's #1693 answers of 3 November 2026 (answers), which do not say this in those words. The craftsman course differs under an owner-approved answer: "The novice craftsman label starts with the craftsman course" (docs/GDD.md:372, #1698 follow-up; R-CRAFT-003). Either may hold a job, which is released in the same commit that starts the course (a reading of docs/gdd/cities.md:42, "Reassignment releases old capacity atomically.", with docs/GDD.md:372, Q43(b) and (c), "can be assigned from the next planning, or at once on cancellation", and docs/gdd/training.md:66, people "return at once"); a teacher teaching elsewhere, or a pupil on a course, is refused, since only a cancel or a graduation ends those here. No reception progress is read (docs/GDD.md:374, "Refugees cannot train until they become civilians"), and refugees in a reception camp take no course place. No tools are needed or charged: tools are post-MVP (#1693, owner answers of 3 November 2026, answer 1: "Tools in the MVP: no."), and the record's tools fields are kept unchanged. A teacher also takes at most Agreed profession.trainees_per_teacher pupils (currently 1; the approved Q42 answer names this value as its basis), so a busy teacher refuses another pupil; a synthetic value above 1 admits no second pupil, because the centre's own limit is one. No Money is charged. Teacher and trainee then hold exclusive Teaching and Course assignments, so other assignment paths refuse them (docs/gdd/training.md:31 at 605277dfa511dc512562ececeb273b8393b21d35, "The teacher is fully occupied and cannot work, fight, build, travel or teach another pupil simultaneously."). Every refusal leaves the state unchanged. Craftsmen, soldiers and refugees cannot enrol. Only exact gdd-0.3 states with an initialized population ledger are accepted; legacy and gdd-0.1 are refused unchanged. The centre course does not retire the Hakhshara: the owner kept it under gdd-0.3 (4 November 2026: "Keep the Hakhshara", #2484), and R-TRAIN-001's halutz track stays.

R-TRAIN-006: In the same fixture, the enrolment month is course turn 1 (#1668 question 43(a); GDD, 1 November 2026 answers: "A course occupies its people from its first turn"), so a course turn is counted in the enrolment month, even with a snapshot taken before the course started, unless the centre was not available at that month's freeze, and in a later month only when its opening snapshot (the #1716 synthetic month context) already holds that course with the same teacher. It is counted at most once per month ordinal: an ordinal before the enrolment month, or not above the last one counted for that pupil, is refused. A counted turn keeps the same population id and every record field (work XP, reception progress, equipment, location, military history); no person is created, and teaching and the course earn no work XP (docs/gdd/training.md:17 at 605277dfa511dc512562ececeb273b8393b21d35, "Travel, idle time, teaching and courses do not count."). The turn that brings a course to Agreed profession.course_turns counted turns (currently 3) graduates it: a novice takes the trained tier through R-QUAL-008's transform; a pupil with no profession takes the course teacher's profession at the trained tier here, not through that transform, with work XP 0 (#2464; the ledger requires work XP 0 with no profession, PopulationLedger.cs:212). The pupil keeps the same id, work XP, reception progress, equipment and the tools' payment fact, location and military history; the pupil and the teacher (a centre has one pupil, R-TRAIN-005) are released in that turn, in one change, as civilians without a job. Nothing is charged and no person is created. #1697: "A free civilian or a matching equipped novice may enrol; completion gives the trained tier." Release timing follows R-TRAIN-008. The fixture also refuses a unit already trained, because Amir's proposal made a trained profession permanent (docs/gdd/training.md:13 at the same commit, "Once trained, a civilian profession cannot be replaced."). That lock is post-MVP: D3 keeps two-way professions (R-QUAL-006), and the GDD's 1 November 2026 answers say "Graduates join their city's civilians and may re-enrol at any time" (27M-5).

R-TRAIN-007: In the same fixture, cancelling a course at any turn before its graduation keeps the population id, profession, novice tier (a pupil with no profession stays without one, #2464), the record's tools fields and reception progress. It grants no unfinished trained tier, clears the course progress and releases the teacher and the centre's place. Nothing is refunded and no one moves. The released teacher and pupil are civilians without a job at once, in that month's snapshot, and may take a job that month, or start another course that month, and the centre can take another teacher and pupil at once (#1716; #1668 question 43(b), owner-approved answer of 1 November 2026, answers: "Cancellation: they are released at once, as civilians without a job."). A second cancel, a cancel of a unit not in a course, or a cancel or further course turn for a graduate, is refused unchanged, and a restart is a full new course with no charge (docs/gdd/training.md:49@605277dfa511dc512562ececeb273b8393b21d35, "Cancelled courses retain paid entry costs and earned reception credit but award no unfinished qualification." and "Unfinished course progress clears, and restarting needs a fresh course."). A centre that stops operating cancels its course at once in the same way (owner, 4 November 2026, https://github.com/eyalgolan/statecraft-israel/issues/1697#issuecomment-6622207214, asked "When a training centre stops operating, what happens to the course running there?", answered "Cancel at once (Recommended)"; the chosen option read: "Losing the centre cancels the course immediately, exactly like CancelTraining: people go home as jobless civilians, no Money refund, issued Arms go back to the pool."; with 27M-4, "It is cancelled like CancelTraining, with no Money refund. Arms return to the pool.", and 27M-9, "To their city as a civilian without a job. Overflow rules apply, with no hard cap.", in the answers): its next course turn counts nothing and cancels the course, the turn that would have graduated it included, so no one graduates, and a cancel there is an ordinary cancel. Our readings, not owner rulings: the fixture's centres are fixed records, so a centre that stops operating is one given to the fixture as not operating; the pupil and teacher are already in the centre's city (R-TRAIN-005), so no one moves; the fixture charges no Money and issues no Arms, so there is nothing to refund or return; and the released pair may start a course that month at another operating centre, while the lost centre refuses a start.

R-TRAIN-008: The fixture refuses, with a "Deferred:" reason and the state unchanged, every case this slice has not built, and chooses no default. Question 44 and the novice lock-in below are decided (GDD, 1 November 2026 answers, and D3), so those refusals are implementation gaps, not open design answers. Question 42 is built and refuses with a plain reason that cites it, not as Deferred (R-TRAIN-005): a second teacher or a second pupil at one centre, a teacher outside the centre's region (the region standing in for the city), and a teaching teacher with no pupil (an invalid inventory); a pupil outside the region is refused with a plain reason that cites the owner's Q23 ruling (4 November 2026) on where the pupil lives, and our reading of 27M-7 for training there with no travel, as R-TRAIN-005 says. Question 43 is built in full: what either participant held at the month's opening (a job, or a course or teaching cancelled this turn) is no bar to starting a course that month (a reading of docs/gdd/cities.md:42, "Reassignment releases old capacity atomically."; docs/GDD.md:372, "they are released in the turn its last course turn counts and can be assigned from the next planning, or at once on cancellation"; docs/gdd/training.md:66, people "return at once"), except a graduation, below. The ruled question 43 limits are refused unchanged with a plain reason that cites the ruling, not as Deferred: counting a month after the enrolment month whose opening snapshot does not hold the course with the same teacher (43(a)). There are no reception beds, so a centre's place freed by a cancel or a graduation is free at once, also in the month a course started and was cancelled; #1668 question 41's bed reuse is post-MVP (docs/GDD.md:377). #1668 question 44 and the GDD's novice lock-in decision: a novice of another profession entering a course (a free civilian without a profession enters, R-TRAIN-005, #2464). A course turn in a month whose #1716 facility freeze lists the centre as unavailable or without a place: a centre that stops operating cancels its course (R-TRAIN-007), but when a loss inside a month takes effect against that freeze is unruled. Graduation release under question 43(b) and (c) is built and is not Deferred (#1697 body, owner answers of 1 November 2026: "on graduation the people are released in the turn the last course turn counts and can be assigned from the next turn's planning" and "a person whose course ends this turn starts work the next turn"): a graduate and a teacher released by the graduation cannot start a course in the graduation month, even after taking a job that month (refused unchanged with a plain reason, not Deferred), and can from the next month; and work eligibility is read from the month's opening snapshot (#1716), where the graduate still holds the course, so a job given after graduation counts from the next month. Real centres, hosts, regions, map bindings, Money and Food settlement, saves and commands remain held with their owners (#1691, #1696, #1700, #1702, #1713, #1716; #1690 closed on 2 November 2026, its post-MVP regional pools moved to #2233).

Craftsmen

R-CRAFT-001: A civilian becomes a craftsman by training (D-D4; Q-CRAFT-1, owner, 15 October 2026): Train(city, craftsman, n) follows R-TRAIN-001 to R-TRAIN-003, and the city's facility is a completed industry project (R-PROJ-003); one still being built does not count. unit_types.yaml gives the course's training_turns (3) and training_cost (IL 0.5m a group), labelled design values (Q-CRAFT-3). A finished cohort adds to the city's count of craftsmen, which is not part of its civilians and is written into the state hash only when not 0. A craftsman can return to civilian life and keep the qualification (R-QUAL-002). Under gdd-0.3, the craftsman course on the population ledger is R-CRAFT-003 (fixture only). A gdd-0.3 game started without the dark start (the public Simulation.NewGame) still runs this aggregate course, the gdd-0.1 behaviour until #1689's last slice (#2528).

R-CRAFT-002: A craftsman fills an ordinary job slot (Q-CRAFT-2, owner, 15 October 2026). Of a city's employed, its craftsmen are counted first, and each employed craftsman earns craftsman_income_multiple (2, a labelled design value) times income_per_job_per_year; every other employed unit earns the R-CITY-002 rate. Under gdd-0.3, a craftsman is a novice from the course start, then an apprentice or an expert (R-CRAFT-003); the course's Industry slot and income stay with #1692. Work XP under gdd-0.3, in synthetic fixtures only (#2472, Gdd03CraftWorkFixture; nothing in the game calls it): docs/gdd/training.md records "Apprentices become experts after craft.expert_work_turns actual turns in Industry or an arms workshop. Each turn counts once, regardless of output/work rate." Each apprentice or expert craftsman the month's snapshot (#1716) has working an Industry or arms-workshop slot gives one work receipt for that population id and month; idle, travel, course, teaching, service, farming and building months give none, and a novice craftsman's workshop month changes no qualification (Q55). Receipts are deduplicated by id and month, within a batch and against those already applied, before any XP, so a repeated receipt adds nothing and cannot promote twice. Each receipt adds one tick whatever the output or pay, and an apprentice whose tick reaches craft.expert_work_turns (36) becomes an expert: the month's snapshot keeps the apprentice tier for that month's output, and the expert tier counts from the next snapshot. Only the tier and work XP change; the id, assignment, location, equipment and military history stay. An expert's work XP keeps counting past 36 (maintainer reading: Gdd03CraftArithmetic.WorkTick already counts it, and no ruling stops it). The detached Gdd03CraftArithmetic no longer counts a building turn as work XP; it still quotes the builder's work.

R-CRAFT-003: In gdd-0.3 synthetic fixtures only (#1698, Gdd03CraftCourseFixture; nothing in the game calls it), a craftsman course on the #1689 population ledger. A free civilian with no profession, whatever their reception progress, which is not read (docs/GDD.md:374: "Refugees cannot train until they become civilians"), or a novice farmer or pioneer, starts the course and becomes a novice craftsman under the same population id, as a trainee on the course (#1698 follow-up and 27M-3, in the 1 November 2026 answers the owner approved, answers: "When the craftsman course starts." and "The course replaces the novice profession, and paid tools are not refunded."). Work XP restarts at 0 because the course replaces the novice profession, a maintainer call; tools and their paid fact are kept. A person with a job may start: the job is released in the same change, and what they held at the month's opening is no bar. This is a maintainer reading of docs/gdd/cities.md:42 ("Reassignment releases old capacity atomically.") with docs/GDD.md:372 Q43(b)/(c), not an owner ruling. A person on a course, teaching or in reception, and a refugee are refused. Course turns follow R-TRAIN-006's Q43 timing: the enrolment month is course turn 1, a later month counts only when its opening snapshot holds the course, and each month ordinal counts once. The turn that reaches Agreed craft.course_turns (currently 3) makes an apprentice craftsman with the same id and every other field, released in that turn as a civilian without a job; they cannot start a course that month (Q43(b), R-TRAIN-008), and work starts the next month. Cancelling before that makes an ordinary civilian again: profession and tier none, work XP 0, released at once, same id, tools and their paid fact kept and nothing refunded (#1698 follow-up: "On cancelling before qualifying they are an ordinary civilian again."; 27M-4; Q44). No teacher is read or created and the population count does not change (#1698: "The course adds no mentor population unit"); a teaching record at a craftsman course is refused. The ledger accepts a novice craftsman outside a course too, a maintainer reading of Q55 ("workshop work by a novice craftsman"). A trained or veteran farmer or pioneer may start the course and loses that level, as a qualified civilian a Train order takes does (R-QUAL-006; docs/GDD.md:349, D3): they become a novice craftsman with work XP 0, keep their tools and the paid fact, graduate as an apprentice, and on cancelling are an ordinary civilian, not the old level (owner, 4 November 2026, on "#1698: may a trained or veteran farmer or pioneer start the craftsman course?": "Yes, lose old level (Recommended)"). An apprentice or expert craftsman may take the course again (27M-5, docs/gdd/training.md:66: "graduates may re-enrol at any time") and keeps their tier on the course, after graduating and after cancelling; the course never lowers a qualification (owner, 4 November 2026, on "#1698: what level does a craftsman hold when they take the craftsman course again?": "Keep their level (Recommended)"). That they keep their work XP too is a maintainer reading, not the owner's: the recorded answer covers only the level. Not built here, refused with a "Deferred:" reason and the state unchanged: a novice craftsman outside a course starting it (the 4 November 2026 answer names an apprentice or expert), and a course turn when the course is not available at the month's freeze (Q52). Out of this rule, with #1692: the Industry slot, its capacity and the course income; and with their owners: the 0.5 Money course cost, apprentice-to-expert work turns (fixture only, R-CRAFT-002), live Train orders and saves. Every refusal leaves the state unchanged; legacy and gdd-0.1 states are refused.

Sites

Construction is step 5 of the turn in the approved order (§2.3, D-A6): after training, before production. The designated sites are sourced facts in content/sites.yaml; Phase 1 has one, Mefalsim (owner decision on #378). The answers below are the plan's recommendations, which the owner took on 14 October 2026 (#378).

R-SITE-001: PioneerWork(unit, build_yishuv, site) starts building a yishuv at a designated site whose kinds include yishuv and which is neither being built nor built, with a halutz or veteran halutz unit that has no other project (D-D6: one project per unit). The Treasury pays build_yishuv_cost at once. When the Treasury holds less, the shortfall is borrowed, as for map actions: it is added to the debt and the Treasury stops at 0 (owner decision "Allow borrowing", 15 October 2026). The site has build_yishuv_turns to go, counting the turn of the order. The unit does not move. It is refused, with the state unchanged and the Attention refunded, for an unknown unit, a unit that is not a halutz or veteran halutz or is busy, an unknown action or site, a site that is not for a yishuv, or a site being built or built (one unit per site). The legacy loop refuses it before charging Attention. Its Attention cost is 0 (D-A6: unit orders are free).

R-SITE-002: StopWork(unit) ends the unit's build: the site is available again, the unit is idle, and the progress and the Money paid are lost (D-A3). A unit that does not exist or is not building is refused with the state unchanged. Its Attention cost is 0.

R-SITE-003: Each turn, in ordinal site id order, every site being built has one turn of work less to go - progress x m of one, the system's office multiplier m since the ministers (R-MIN-002): the integer part at once, the fraction carried on the record and written after it (as 16 hexadecimal digits of its IEEE bit pattern) only while non-zero. A site that reaches 0 is built, on the turn's end date, and its unit is idle again. A veteran's site has veteran_progress times as much work a turn (R-VET-002), and the unit's type changes only by R-VET-001. No city, housing, job, population or Social cohesion changes. Sites being built or built are written into the state hash only when there are some.

Veteran pioneers

R-VET-001: Under legacy/gdd-0.1, and under gdd-0.3 until the population ledger's work-XP tiers are live, in step 5, after its site's or segment's progress (R-COR-002), a halutz unit whose site or corridor segment is being built gains one turn of service, the turn it is built included; an idle unit gains none. At veteran_halutz_service turns of service it becomes a veteran_halutz at the end of that turn, keeping its id, hex, service and project, and a cause says so. A veteran gains no more service. A unit's service is written into the state hash only when it is not 0 (D-D9; Q-VET-1, owner, 17 October 2026, #782). A halutz or veteran can return to civilian life and keep its service or its status (R-QUAL-003). In gdd-0.3 synthetic fixtures only, the pure profession helper proposes one actual-work XP increment for a received free civilian with a matching Work assignment and completed staffed-farming evidence regardless of output or advancement, or evidence of advancing an on-site pioneer project. Adopted sources at 605277dfa511dc512562ececeb273b8393b21d35: docs/gdd/training.md:17, "Farmers earn work XP only on staffed farming turns. Pioneers earn it only while advancing an on-site pioneer project."; docs/gdd/balance.md:70, "Work XP counts actual qualifying turns, not output or work quantity." Multiple matching results count once per evaluation; zero pioneer advancement, wrong profession/target, generic work, travel, idle, teaching and courses give none. Both novice and trained farmer/pioneer become veteran at the Agreed catalogue profession.work_to_veteran threshold (currently 9). Craftsmen do not use this path. The caller must use one turn-start snapshot and commit at most once at turn end; turn scheduling, authoritative evidence production and cross-call deduplication remain held. No game-loop caller is added. The fixture-only pioneer producer (Gdd03PioneerWorkFixture.ResolveTurnEnd, #1693) is that caller for pioneers on initialized gdd-0.3 ledger fixtures: from one turn-start state and the pioneer projects being built that turn (yishuv sites and corridor segments alike; owner, 3 November 2026, #1693 answer 5), each pioneer record at Work on one of them gives advancement evidence and its result is committed once, in ascending id, into a new state, except that a veteran's work XP is not advanced (maintainer call, not a ruling, matching the legacy veteran that gains no more service; it still works at its rate); the turn-start state is unchanged, so resolving the same snapshot twice gives one tick. Records off the job (idle, travel, course, teaching) earn nothing and keep their tier and work XP (answer 4). MVP pioneers are course-trained (answer 2). The fixture-only farmer producer (Gdd03FieldFixture.ResolveWorkXp, #2468) is the same caller for farmers: from one turn-start state and the month's snapshot (27M-2), each farmer the snapshot has working a field that is still Israeli-held (Q37, the fields R-FIELD-002 pays) gives one staffed-farming receipt, committed the same way, with the same veteran exception; a farmer whose turn-start record is no longer on that field, idle, travelling, teaching or in a course earns nothing. It has no game-loop caller until #1689's integration slice. The ledger harvest (R-FIELD-002, #1694, #1725 PR 7) commits no work XP either (owner, 8 November 2026, on #1725, "Not in this change (Recommended)").

R-VET-002: Under legacy/gdd-0.1, and under gdd-0.3 until the population ledger's work-XP tiers are live, a veteran's site or corridor segment being built has veteran_progress turns less to go each turn, against a halutz's 1; one with that many or fewer left is built that turn. A veteran does the same pioneer work as a halutz (R-SITE-001) (D-D9; Q-VET-2, owner, 17 October 2026, #782). In gdd-0.3 synthetic fixtures only, output quotes read the turn-start tier from Agreed catalogue pioneer.work_rate (novice/trained/veteran: 0.75/1/1.5) or farmer.output (2/3/5, multiples of field_food_per_turn, R-FIELD-002 (owner, 1 November 2026: "Main's base x 0.3's tiers (Recommended)"; #2112); the farmer quote here and in the synthetic field evaluator is that multiple, which equals Food only while field_food_per_turn is 1, and Gdd03FieldRates.FoodPerTurn converts it). Promotion returns a separate record for the next turn, preserving identity, assignment, placement, equipment and independent military history. No project progress or farming yield is applied by the helper; absent, Deferred or malformed values fail closed without legacy fallback. The #1693 pioneer producer quotes each record's work with this turn-start pioneer.work_rate, never veteran_progress, and returns the quote per target, summed over the records at Work on an advancing target, without applying it to any project. The ledger harvest (R-FIELD-002, #1694, #1725 PR 7) applies the farmer quote: a field that yields adds farmer.output at its month-snapshot farmer's tier times field_food_per_turn to the Food stock.

Return to civilian life and reassignment (#1511)

Owner ruling "Two-way, keep training" and its three answers (23 October 2026, #1511): https://github.com/eyalgolan/statecraft-israel/issues/1511#issuecomment-6372356482 and https://github.com/eyalgolan/statecraft-israel/issues/1511#issuecomment-6372642736. Craftsmen, halutzim and veteran pioneers can go back to civilian life and be put back into the trained role, in either direction, without repeating R-TRAIN-001.

R-QUAL-001: Each city keeps its qualified civilians (City.Qualified): qualified craftsmen, qualified halutz groups (each with the turns of service it had, ascending) and qualified veteran pioneers. They are part of the city's civilians, not in addition to them, so a qualified civilian is an ordinary civilian for jobs (R-CITY-001, R-CITY-009), housing (R-HOUSE-003), income (R-CITY-002, the ordinary rate, not R-CRAFT-002's) and training (R-TRAIN-001), and keeps the qualification for reassignment (owner ruling 2). The sum of the three never exceeds the city's civilians.

R-QUAL-002: ReturnCraftsmen(city, n) turns n of the city's craftsmen into civilians and qualified craftsmen, at once and with no Money. Employed is unchanged. It is refused, with the state unchanged and the Attention refunded, when the city is unknown, n is below 1 or n is above the city's craftsmen. On a game with an initialized Historical ledger (the dark start, R-POP-002; #1725) the order is refused, after the unknown city and the count below 1, with "Craftsmen are not on the population ledger yet (#2476)", the state unchanged and nothing thrown, until #2476 makes craftsmen exist there: the Historical start has none, and the craftsman course is refused until then (R-TRAIN-001) (owner, 9 November 2026: "Refuse for now (Recommended)"). A synthetic or Open ledger is still refused (R-POP-002). This internal English text gets its Hebrew with the UI wiring PR.

R-QUAL-003: UnitOrder(unit, return_civilian) (the unit card's "Return", an order with no hex) takes an idle halutz or veteran pioneer unit off the map, at once and with no Money. It joins the nearest city (hex distance to the settlement hex, ties by ordinal city id, R-UNIT-005's rule) as one civilian and one qualified halutz group with the unit's service, or one qualified veteran pioneer. A halutz's partial service toward veteran is kept (owner ruling 3). It is refused, state unchanged, for an unknown unit, a unit that is not a halutz or veteran ("Only a pioneer unit takes that order") and a unit at work on a site or corridor segment ("That unit is at work; stop the work first"). On a Historical ledger game (#1725 PR 7; owner, 9 November 2026: "Add the link first (Recommended)") exactly the unit's record returns. "Nearest city, settled (Recommended)": it becomes a free civilian without an assignment in the nearest city (its region plus the city), with reception ticks equal to reception.turns, as Send home does (R-UNIT-005), and keeps its profession, tier, work XP and items, except that a veteran pioneer unit's record becomes a Veteran pioneer (R-QUAL-005). "City counts, as today (Recommended)": the unit's service, or the veteran, goes into the city's City.Qualified as above, and no new field is added to the record; the civilian count is the records', so no other city field is written. A pioneer unit with no linked records (one an event grants) makes new ones (owner, 9 November 2026: "Make new records"): "One per unit (Recommended)", one new person record, registered under its "Own heading (Recommended)", returned_pioneer_unit/<city> with ordinals from 0, apart from the census, the natural increase and the arrivals, so "Yes, new people (Recommended)": it counts as a new person in the ledger's totals, as an arrival does. "Beginner; veteran stays (Recommended)": a halutz gives a Novice pioneer with 0 work XP and a veteran pioneer a Veteran pioneer. It lands as the other returned pioneers do, free, unassigned and settled in, in the nearest city, with no qualification, combat XP or item, and the service goes to City.Qualified as for any return.

R-QUAL-004: ReassignCraftsmen(city, n) turns n of the city's qualified craftsmen, who have no job, into craftsmen again, at once, with no Money and no training. It is refused, state unchanged, when the city is unknown, n is below 1, n is above its qualified craftsmen or n is above its civilians without work (R-TRAIN-001's Idle; owner, 24 October 2026: "Accept all as built", https://github.com/eyalgolan/statecraft-israel/issues/1511#issuecomment-6375597377: the check counts all the city's idle civilians). Queued, it counts in full against the city's civilians without work as a queued Train does (R-TRAIN-001), so the orders are accepted in any order, and it counts against Cities.MaxWorkers (R-CITY-001) for the turn it is queued: its craftsmen stay in the city, but if the job order runs first they are already in work. On a game with an initialized Historical ledger it is refused as R-QUAL-002 sets, with the same reason (owner, 9 November 2026: "Refuse for now (Recommended)").

R-QUAL-005: ReassignPioneers(city, track, n) (track is halutz or veteran_halutz) turns n of the city's qualified halutz groups or veteran pioneers, who have no job, into n units at the city's settlement hex, as R-TRAIN-003 places them, taking the next ids in turn: a halutz keeps its service and a veteran comes back a veteran with veteran_halutz_service, and the halutz groups with the most service come back first (owner, 24 October 2026: "Accept all as built", https://github.com/eyalgolan/statecraft-israel/issues/1511#issuecomment-6375597377). Refusals as R-QUAL-004's, and "There is no such qualification" for another track. Queued, it counts against the city's civilians without work as a queued Train does (R-TRAIN-001) and against Cities.MaxWorkers (R-CITY-001), since its groups leave the city as units. On a Historical ledger game (#1725 PR 7; owner, 9 November 2026: "Free pioneers by service (Recommended)") the groups are the city's free, unassigned pioneer records in ascending id, paired with the qualified groups most service first, so the group with the most service is the lowest id; each record goes back to region-only with no city, as at graduation (R-TRAIN-003), and its new unit is linked to it. A returned veteran's record counts for the veteran_halutz track ("Yes, keep the link (Recommended)"), and following that ruling there is no split between the tracks, only a preference: each track takes its own tier's records first (Veteran-tier ones for veteran_halutz, the others for halutz), then the other tier's when it has too few, so no qualified group is left without a record it can use; a linked veteran pioneer unit's record returns as a Veteran pioneer (R-QUAL-003), as a new record from one does. The training centre's census veteran pioneer teacher, the last census/haifa record (#1689), is set apart only while untouched, never linked to a map unit since the start: it then comes after every other record, of either tier (owner, 7 November 2026: "Takeable, but last (Recommended)"), and a Train order does not count it as holding a qualified group (R-QUAL-006). Once a unit is linked to it, the ledger records that (teacher-linked, in the state hash only once set) and it is a pioneer like any other. The order is refused, state unchanged, when the city has fewer free, unassigned pioneer records than n, as when a qualified pioneer holds a job (an arms workshop's worker, R-ARMS-005), the ledger's form of "civilians without work"; only City.Qualified is written.

R-QUAL-006: A Train order (R-TRAIN-001) takes the city's unqualified civilians first. A qualified civilian it has to take loses its qualification, in the order craftsmen, halutz groups with the least service, veterans (owner, 24 October 2026: "Accept all as built", https://github.com/eyalgolan/statecraft-israel/issues/1511#issuecomment-6375597377). On a Historical ledger game (#1725 PR 7; owner, 9 November 2026: "Yes, lower it (Recommended)") a Train order that takes a returned pioneer record lowers the city's qualified pioneers (City.Qualified) by one, halutz groups with the least service first, then veterans, so the qualified pioneers never outnumber the city's free civilian pioneer records; as on the aggregate path, only what the remaining records cannot hold is lost ("Only as needed (Recommended)").

R-QUAL-007: Attention is that of the matching existing order (owner ruling 1): return_craftsmen and reassign_craftsmen cost what assign_workers does (0.5), and reassign_pioneers and the return_civilian unit order cost what a unit order does (0), all design values in rules.yaml. The qualified counts are written into the state hash only when not 0 (,qcraftsmen=N,qhalutz=S:S,qveterans=N after the other city fields), so a game without them hashes as before. The saves write retcraft CITY N, regcraft CITY N, regpioneer CITY TRACK N and order UNIT return_civilian; scenarios use return_craftsmen, reassign_craftsmen, reassign_pioneers and unit_order with order: return_civilian.

R-QUAL-008: Region-independent gdd-0.3 fixture helpers only (coordinator decision, 26 Oct 2026, Q5): a matching farmer/pioneer course may propose trained tier while retaining actual-work XP, veteran tier and every other record field. Cross-profession qualification is refused without mutation. This transform does not decide course eligibility, completion or release. Tools quotes read only Agreed profession.farmer_tools / profession.pioneer_tools (currently 3 / 5); no acquisition, charge, credit, refund or paid-flag change occurs. All helpers are internal and require an initialized exact-gdd-0.3 fixture. Money policy, once-only acquisition, lock-in policy, craftsman progression and production integration remain held with their owners. Tools stay a quote: the owner ruled the one-time tools charge post-MVP (3 November 2026, #1693 answer 1).

R-SITE-004: Every designated site has one or more kinds from a fixed list: yishuv, field, swamp, military_camp and hakhshara (Phase 4, Q-FIELD-1, Q-SWAMP-1, Q-CAMP-3). A pioneer action is accepted only at a site of its kind; any other is refused with the state unchanged and the Attention refunded. A yishuv site has a founding date inside the MVP's years and no other kind has one. A camp site (military_camp or hakhshara) names a city of cities.yaml, the city it trains for (Q-CAMP-1), and no other kind names one. A yishuv site is never also a camp site (military_camp or hakhshara): the two would share one built record, so building the yishuv would count as a built camp for its city (#2599). The loader refuses that row and any other invalid one, naming the file and field.

R-SITE-005: At most one yishuv site sits in a hex, so a hex shows at most one new yishuv. A site of another kind may share a hex with a yishuv site, as a yishuv's field site does (Q-SITE-3).

R-FIELD-001: PioneerWork(unit, plow_fields, site) plows a field at a site whose kinds include field, by R-SITE-001 to R-SITE-003's rules, with plow_fields_turns and plow_fields_cost (labelled design values, Q-FIELD-3, decided 17 October 2026). A field site in a yishuv site's hex is ready only once that yishuv is built; before that the order is refused with the state unchanged and the Attention refunded (Q-FIELD-1, decided 17 October 2026). Built, the field site is plowed. A field site is not also a yishuv site, so building a yishuv cannot plow it; the loader refuses a row that gives it both kinds.

R-FIELD-002: Each turn in step 6, after the cities' production and in ordinal site id order, every plowed field adds field_food_per_turn Food units to the national stock (R-FOOD-001), with no workers (Q-FIELD-2, decided 17 October 2026). A field being plowed adds nothing. The Food arrives before the nation eats in step 7, so in a turn with imports (R-FOOD-002) the import tops the stock up by that much less. field_food_per_turn is 1, a labelled design value (Q-FIELD-3: "1 Food a turn"), and required in the GDD loop; a value below 0 or above 1,000 (Pools.MaxFieldFoodPerTurn, #2804) is a load error. A gdd-0.3 game started without the dark start (the public Simulation.NewGame) still runs this aggregate harvest, the gdd-0.1 behaviour until #1689's last slice (#2528).

Then changed (3 Nov 2026): under exact gdd-0.3, in synthetic fixtures only (R-POP-004), each authored starting field (regions.yaml field_hexes, Q36) is one Work slot for one farmer (field.slots 1; owner, 3 November 2026, "One farmer per field (Recommended)"), staffed only by a farmer located in its region (Q27) while the field is Israeli-held (Q37). Only a farmer may hold a field (owner, 4 November 2026: asked "May only a unit with the Farmer profession staff a field?", the owner answered "yes", recorded on #1694). A field staffed in the month's snapshot (27M-2) and still Israeli-held yields farmer.output (2 / 3 / 5) times field_food_per_turn by the snapshot tier, and no Money; an unworked field yields nothing. Then changed (#2471): the fixture's settlement adds that yield to the national Food stock (R-FOOD-001), each field's once and in ordinal field-id order, with one Why line per field that yields. Control is checked again at settlement as well as at the snapshot's freeze, and a field lost by either adds nothing while its farmer keeps the same id, tier and assignment (D6: "its people stay"); once retaken, the next month's snapshot yields again with no new assignment (D6: "On retaking, everything resumes and nothing is deleted."). The caller states which fields are Israeli-held (maintainer reading: the sim has no hex control for fields yet). No game-loop caller is added, and the legacy and gdd-0.1 plowed-field harvest above is unchanged.

Then changed (#1694, #1725 PR 7): on a game with an initialized Historical ledger (the dark start, R-POP-004), the step-6 harvest (Pools.Harvest given a month) reads the turn's one month snapshot (Gdd03SyntheticMonthContext.CreateLive; owner, 8 November 2026, on #1725, "One snapshot per turn (Recommended)": the snapshot is taken at the turn's start, after the orders). In ordinal field-id order, each authored field yields farmer.output at its snapshot farmer's tier times field_food_per_turn only if that farmer's record still exists and still holds Work on the field when the harvest runs ("Yes, recheck (Recommended)"), and the field is held: no enemy military unit stands on its hex at that time, the test Cities.Held uses for cities (R-CITY-010; "Held unless enemy on it (Recommended)"). An unworked or enemy-held field adds nothing, and its farmer's record is unchanged. Each field that yields writes one field food Why line (field_food, its argument the field id; "Per field, words later (Recommended)"), whose EN and HE wording comes with the dark-turn change. A plowed sites.yaml field adds nothing on a ledger game until #1713 ("Nothing until #1713 (Recommended)"). The harvest gives no work XP (R-VET-001). Every check, the checked sum included, comes before the stock changes. No turn passes it a month yet; without one the harvest is the aggregate one above, still refused on any initialized ledger (R-POP-002).

R-SWAMP-001: PioneerWork(unit, dry_swamp, site) drains a swamp at a site whose kinds include swamp, by R-SITE-001 to R-SITE-003's rules, with dry_swamp_turns and dry_swamp_cost (12 turns and IL 4m, labelled design values, required in the GDD loop; Q-SWAMP-1, decided 17 October 2026). Built, the swamp site is drained, and the field site in its hex becomes ready: before that a plow_fields order there is refused with the state unchanged and the Attention refunded, as at an unbuilt yishuv (R-FIELD-001). Drained, the swamp does nothing else. A field site is not also a swamp site, since draining would plow it; the loader refuses a row that gives it both kinds. The MVP's one swamp site is the Hula (Q-SWAMP-1; Q-SWAMP-3, "Allow as planned", owner, 17 October 2026, #789).

R-SWAMP-002: An event with site: <id> names a swamp site (the loader checks it in the GDD loop) and, in the GDD loop, can fire only while that site is being worked: it has a record that is not built. With no works, or once the swamp is drained, it waits, and it can still fire later in its window (Q-SWAMP-2, decided 17 October 2026). The legacy loop reads no sites and ignores the key, so the legacy goldens hold. hula_drainage carries site: hula.

Camps

R-CAMP-001: PioneerWork(unit, build_camp, site) builds a training camp at a site whose kinds include military_camp or hakhshara, by R-SITE-001 to R-SITE-003's rules, with build_camp_turns and build_camp_cost (labelled design values, required in the GDD loop). One action builds either facility, as the site's kind says (Q-CAMP-3, owner, 17 October 2026, #782). Built, on the turn's end date, the camp gives the site's city (Q-CAMP-1) the site's facility, with no capacity limit under live gdd-0.1, so Train accepts that city for the matching track from the next turn (R-TRAIN-001). Under gdd-0.3, 27M-25 says: "every camp uses military.camp_capacity 5" (50 units at D2's 1,000 people a unit, content/gdd03-balance.yaml:307). A camp being built gives nothing. The facility is read from the built site's record, so the state hash gains nothing beyond the site's own record. The MVP's camp sites are Camp 80 (for Hadera) and Tzrifin (for Rishon LeZion), both military_camp (Q-CAMP-2, owner, 17 October 2026, #785; "Both military", owner, 17 October 2026, #927).

Then changed (1 Nov 2026, #2112): the owner chose the option labelled "Main's base x 0.3's tiers (Recommended)", so counts sized in 10,000-people units are 10x at 1,000 people a unit: military.camp_capacity is 50 units, as content/gdd03-balance.yaml stores it.

R-CAMP-002: A camp site's city is the MVP city (cities.yaml) nearest to the site by hex distance (Q-CAMP-1). When several are equally near, it is the one of them nearest by straight-line km from the point the site's hex comes from to the city's point in map/settlements_1948.yaml (owner, 17 October 2026, #927). Tzrifin's hex is one hex from Tel Aviv, Petah Tikva and Rishon LeZion (4.4 km hexes, #1183), so it serves Rishon LeZion, about 3.9 km away.

R-CAMP-003: Under gdd-0.3 only, Tel Aviv starts with the existing Tel HaShomer camp and Jerusalem starts with Schneller. The owner's 3 November 2026 ruling is "Tel Aviv's camp is Tel HaShomer; add Jerusalem" (recorded on #1719). gdd03-start.yaml names the map location each start camp stands for: military_tel_litvinsky for Tel Aviv and military_schneller for Jerusalem. The loader keeps no link from a city to its location: only the city's military_camp facility flag is kept, added when the city lacks it, and Training checks that flag (R-TRAIN-001). These choices follow the ruling, not R-CAMP-002. When content/gdd03-start.yaml is present, the loader requires unique cities and unique military camp locations, each location must be an existing military_camp map location, and the effective number of cities with camps must equal start.military_camps. The file is optional, so synthetic gdd-0.3 fixtures load without it. The shipped file is pinned by tests: deleting it fails 4 tests, Gdd03StartFacilitiesTests.ShippedGdd03StartHasTheCataloguedNumberOfCampCities, Gdd03StartFacilitiesTests.Gdd03ReadsOverlayBytesIntoItsHash and the two SchemaTests cases for gdd03-start.yaml (RealContentPassesItsSchema and RealContentNamesItsSchemaInTheFirstLine). Under gdd-0.1, the overlay is not read and Tel Aviv keeps the one start camp of D-D5.

Military units

Phase 5 PR 2 (#741) introduced the military record and types. PR 3 (#742, #1130) adds recruit formations, service and Money upkeep. Phase 5 PR 4 (#743) adds the Arab forces' types and the sourced starting forces (R-UNIT-006).

R-UNIT-001: A military unit is a group at formation scale, one token per formation (D-A1; Q-UNIT-1). It has a side, a type, a hex, a strength (an integer, in population units) and a standing order with its target. A military type in unit_types.yaml is a unit with a side (one of units_1948.yaml's sides) and a class (infantry, artillery or armour); the loader refuses a type with only one of them, an unknown side or class, or a population type with either, naming the file and field. The MVP's military types are Israel's recruit, veteran_soldier, artillery and armoured_car.

R-UNIT-002: A unit's side, strength, order and order target are written into the state hash only when they are not at their default ("" or 0), after its service (R-LOOP-002's rule). Its housing home city is also written when nonempty (|home=<city>); an empty home adds nothing, preserving earlier units' canonical text. Its combat experience (R-UNIT-011) is written last, after |full=, only when it is not 0 (|xp=<n>), so no gdd-0.1 unit's text changes. A halutz unit's canonical text remains unchanged when its home is empty.

R-UNIT-003: A finished recruit cohort becomes exactly one Israeli infantry formation at its city's settlement hex, with Strength equal to the cohort's Count in population units, service 0 and the next id (Q-UNIT-1, Q-UNIT-4; owner, 15 October 2026). The cohort leaves the state: nobody is lost or duplicated. Its reserved Arms are neither charged again nor refunded on completion (R-ARMS-003). A cohort aimed at a unit adds to that unit instead of making a new one (R-TRAIN-004).

R-UNIT-004: Each Israeli recruit formation already in play before step 4 gains one service turn; a cohort graduating in step 4 starts serving on the next turn. At veteran_soldier_service it becomes veteran_soldier, keeping id, hex, service, side, strength and standing order, with a Why cause. Veterans stop accumulating service. Combat gives no extra service, and Phase 5 adds no return to civilian life (Q-UNIT-2, Q-UNIT-6; owner, 15 October 2026). The owner-accepted initial threshold is 6 turns, a labelled design value under Q-GEN-1, not a historical period (19 October 2026, https://github.com/eyalgolan/statecraft-israel/pull/1130#issuecomment-6269103829). The loader requires it to be at least 1 and requires the Israeli veteran infantry type when recruits exist. Under exact gdd-0.3 service is a statistic only and never promotes: a recruit still gains its service turn, and veterancy comes from combat experience, R-UNIT-011 (owner, 1 November 2026, Q16: "Under gdd-0.3, yes: veterancy comes from combat experience. gdd-0.1 keeps R-UNIT-004."). Service also runs on a game with an initialized Historical ledger (the dark start, R-POP-002; #1725 PR 7): a recruit formation gains its service turn as in the aggregate game and, under gdd-0.3, never promotes. Service stays on the formation only, so no ledger record or ledger hash changes, and service checks no formation strength against the ledger's soldiers; a synthetic or Open ledger is still refused (maintainer calls by coordinator eyalgolan-30, 8 November 2026, not owner rulings).

R-UNIT-005: UnitOrder(unit, send_home) (the unit card's "Send home", #1309) is the player's order and the only way home: no automatic demobilisation happens (owner, 22 October 2026, #1309: https://github.com/eyalgolan/statecraft-israel/issues/1309#issuecomment-6334162964). It is refused, the state unchanged, with the orders' own reasons for an unknown unit, a unit with no side or an enemy unit (R-ORDER-001's wording) and "That order names no hex" when a hex is given, and with "The armistice with is not signed" while the armistice on the unit's front is unsigned. A unit's front is the front of the enemy side it attacked this turn or borders; ties take the nearest enemy unit, then the lowest unit id (owner, 22 October 2026, #1309: https://github.com/eyalgolan/statecraft-israel/issues/1309#issuecomment-6340226644). A side that belongs to no front (the ALA and the local forces, Q-OBJ-1; R-POW-004's map) names none (accepted per maintainer comment, 22 October 2026, #1309: https://github.com/eyalgolan/statecraft-israel/issues/1309#issuecomment-6341877062: the ALA and the local forces belong to no front, having no armistice of their own), and a unit attacking and bordering no front-bearing enemy may be sent home once any armistice is signed, refused with "No armistice is signed" until then (accepted per maintainer comment 6341877062 on #1309: "no enemy units in play" means attacking and bordering no enemy that has a front). Taken, the order waits for End turn and can be withdrawn until then, as Send home is not an instant order (R-TURN-004); when it runs, it acts at once: the unit leaves the map and its strength, in population units, joins the nearest of the twelve cities as civilians without a job (owner, 22 October 2026, #1309: https://github.com/eyalgolan/statecraft-israel/issues/1309#issuecomment-6334162964: "nearest ... as unemployed civilians"), counting in the city's jobs and housing from the same turn; nearest by hex distance to the city's settlement hex, ties by the city's id (agent default - the ruling sets no tie break). The Why line states the count in people (Q-CIV-4): "Unit N is sent home: its P people are civilians in now". A veteran keeps nothing (none suggested, #1309). On a game with an initialized Historical ledger (#1699, #1725 PR 7; owner, 8 November 2026, #1725: https://github.com/eyalgolan/statecraft-israel/issues/1725#issuecomment-6726421643 and https://github.com/eyalgolan/statecraft-israel/issues/1725#issuecomment-6727007834) the formation's soldiers are person records; a synthetic or Open ledger is still refused (R-POP-002). The refusals above come first. "Every soldier in it (Recommended)": every active-soldier record serving the formation's target (units_1948/<row> for a starting force, formation/<unit id> after graduation, R-TRAIN-003) leaves service. "Refuse with a reason (Recommended)": when the formation's strength differs from those records, the order is refused with Unit N has S soldiers on record but strength T, the state unchanged. "Nearest city, unemployed (Recommended)": each record becomes a free civilian without an assignment, located in the nearest city by the rule above (its region plus the city), so that city houses it from the same turn. "Settled in (Recommended)": its reception ticks are reception.turns. "Keep both (Recommended)": it keeps its military qualification and combat XP. "Back to Arms stock (Recommended)": its issued military item leaves the record and goes back to the Arms stock once, so spare plus issued Arms is unchanged. The opening formations' items were issued apart from the stock (R-POP-004), so sending an opening formation home adds Arms the stock never paid for; this effect was known and accepted with the ruling. "No, records only (Recommended)": no aggregate city field is written, so the city's jobs share (R-CITY-005), which still reads those fields, does not count them. The Why line is the aggregate game's. "Yes, one Why line (Recommended)" (owner, 8 November 2026, relayed by the coordinator for #1725 PR 7): when Arms go back to the stock, one more Why line, after it, names them as training's Cancel does (R-TRAIN-002): ": unit N returned K Arms", its delta the Arms added (key sent_home_arms_returned); none is written when no item goes back, and the aggregate game writes none. Every check comes before the first write; the records change in one validated commit, then the unit leaves and the stock rises. The Arms change is Send home's, not the ledger's.

R-UNIT-006: A new Historical game places one military unit per row of map/units_1948.yaml that is not hidden, has service: army (the squadron and the ships stay presentation-only tokens, Q-UNIT-5), has a strength whose as_of is the start day, and has an anchor window [from, to) covering the start. The unit stands on the hex that holds the first such anchor's coordinates, by the map's projection and cube rounding (tools/hexgrid.py, Unity's HexPicker), precomputed by tools/unit_anchor_hexes.py into map/unit_anchor_hexes.yaml, which the sim reads (no projection maths in the sim, docs/determinism.md); the token may stand one hex away (UnitPlacer), which is presentation. Every IDF brigade is a veteran_soldier; an Arab force takes its side's one military type, a labelled design value of class infantry with 3 movement points. Its strength is the row's people in population units, the nearest unit with a half up and never below 1 (people_per_unit 1,000: Golani's 4,095 is 4, Kiryati's 2,504 is 3). The units take ids 0 upward in the file's order, before any other unit (owner, 20 October 2026, #743: "all five defaults"). The file is in the content hash (Q-START-1), and so is the hex file. The loader refuses a starting row whose strength.value is not a whole number of at least 1 (#2815, option 1, the issue's Recommended option, confirmed on the issue on 6 November 2026), whose anchor is off the map or not on land or has no row in the hex file, a hex row for an anchor the unit file does not have, or whose side has no single military type, naming the file and the row. A shareable build refuses while any row a unit starts from is not fact-checked (Q-START-3): Statecraft.Cli --release-gate, Build -shareable and tools/package_release.py name them.

R-UNIT-007: Open mode starts no forces (Q-START-2, owner, 15 October 2026); Simulation.NewGame refuses Open mode until milestone M4.

R-UNIT-008: Artillery and the armoured car are acquired by training at a military camp (Q-ARTY-3, owner, 15 October 2026, #375: "Take all recommendations"; Phase 5 PR 9, part 2, #748). Each is a Train track (R-TRAIN-001 to R-TRAIN-003): Train(city, artillery, n) takes n groups of the city's civilians without work as the crew, pays n times the type's training_cost in Money and takes n times its training_arms in Arms when it starts, refused for a short stock and returned on cancel as R-ARMS-003 says for recruits. A finished cohort becomes one formation of its type at the city's settlement hex, its strength the cohort's count, housed in that city (R-UNIT-003, R-HOUSE-003, Q-HOUSE-5). It is not a recruit, so service never promotes it (R-UNIT-004); it pays the military upkeep (R-UNIT-010). The loader requires training_arms (at least 1) on a trained military type other than the recruit, whose Arms stay mvp.arms_per_recruit_group, and refuses it on any other type. The values are agent defaults proposed on #748 for the owner to set or accept (https://github.com/eyalgolan/statecraft-israel/issues/748#issuecomment-6341182392): 3 turns, IL 1.0m a group, and 3 Arms a group for the guns, 2 for the armoured car. Event grants of either type stay with Phase 7 (D-H2, R-EFX-003).

R-UNIT-009: A row of map/units_1948.yaml that meets R-UNIT-006 except that its strength's as_of is after the start day arrives on that day (#1241, owner, 21 October 2026, answer 1(a): "enters the game on that date with its sourced strength, read from its row as a dated arrival"): in the GDD loop's resolution (step 9, after the immigrants' arrivals) of the turn that starts before that day and ends on or after it, once, on the hex of the anchor whose window covers that day, with that strength in population units and its side's type, as a starting force. Such rows follow the starting rows in Content.StartingForces, and their unit ids are reserved when the game starts, so unit i is still row i and the release gate names them too. An arrival ignores the stack limit and the hex's other units, as a starting force does. A row whose as_of is before the start starts nothing; the loader refuses an as_of that is not a YYYY-MM-DD date, naming the file and the row. Only a player-side (side: israel) row arrives: the loader refuses an enemy-side row whose as_of is after the start, naming the file and the row, and the resolver refuses to add one (R-ENEMY-004). Today the 7th Brigade arrives: 1,650 on 24 May 1948 at Hulda, 2 units from the 15 June 1948 turn. Under exact gdd-0.3 it also runs on a game with an initialized Historical ledger (R-POP-002, R-POP-004; #1725 PR 7, producer 08), where the arriving brigade's people are ledger records (owner, 8 November 2026, answers to the coordinator): "Soldiers of brigade (Recommended)", so each is an active soldier whose ActiveService target is the brigade (units_1948/<row id>); "New people (Recommended)", so the brigade brings new records, one per strength unit at the ledger's 1,000 people a record (the 7th Brigade's 1,650 men are 2 units, so 2 records), with new ids, none taken from the civilians; "1, like the others", so the 7th Brigade's soldiers start with combat XP 1, its content/gdd03-start.yaml formation_combat_xp row, as the starting brigades' do, and the qualification of its type's class (infantry), as theirs; "Apart from Arms (Recommended)", so each carries one issued military item, units_1948/<row id>/<ordinal>, owned and recorded as paid, and the Arms stock is untouched, as at the start. Coordinator calls, 8 November 2026 (not owner rulings): the dark start binds each own dated brigade at the start (its service target and arrival region, with no record), so the arrival only registers its records; the records are located in the region of the arrival anchor's hex, region-only, as a starting formation's (the 7th Brigade's Hulda hex is in central_plain); and a brigade whose records are already registered adds nothing, so a replay of the turn adds nobody twice. Its records are registered all or nothing before its unit is added; its reception is 0 and it has no profession, as a starting soldier. Its unit is the aggregate game's, with combat XP 0. A synthetic or Open ledger is still refused.

R-UNIT-010: In step 7, after production, Food consumption/imports and civilian upkeep, and before housing, each Israeli military formation pays military_upkeep_per_turn Money, including a formation that graduated this turn. The owner-accepted initial rate is IL 0.05m per formation per turn, independent of strength and turn length, a labelled design value under Q-GEN-1 (19 October 2026, https://github.com/eyalgolan/statecraft-israel/pull/1130#issuecomment-6269103829). Civilian, enemy and default records without a military side pay none. R-MONEY-003 borrows a shortfall; no strength is removed or Arms consumed for upkeep (Q-UNIT-3; owner, 15 October 2026). The loader requires a finite rate >= 0. R-ARMS-003 remains the training/refund rule; the Phase 5 plan's suggested reuse of that id is obsolete. Under exact gdd-0.3 with an initialized Historical population ledger (the dark start, R-POP-004; #1699, #1725 PR 7), the count is unchanged: it counts the formations on the map, each paying the same rate whatever its strength, and reads or writes no soldier record (owner, 8 November 2026, on #1725: "Formations, unchanged (Recommended)"). A synthetic or Open ledger is still refused (R-POP-002).

R-UNIT-011: Under exact gdd-0.3 only, a military unit gains combat experience from the engagements it survives (#1707; owner, 2 November 2026: "Build it now (Recommended)", on today's aggregate units, gdd-0.1 unchanged). After an engagement's one draw, its losses, retreats and surrenders (R-COMBAT-002, R-COMBAT-003), each attacker and defender still in the state, in ascending id, gains the Agreed catalogue military.combat_xp_gain (1), up to military.combat_xp_full (5): both sides, the player's instant attack and step 2's alike, a retreated defender included, and artillery firing alone too (owner, 1 November 2026, #1668 answers https://github.com/eyalgolan/statecraft-israel/pull/1668#issuecomment-6553884547: Q14 "Yes. Artillery fires alone from range, and it earns experience for each engagement it survives, like any unit."; Q15 "It stops."). A destroyed or surrendered unit gains nothing. Moves, orders, service and training give none. A veteran_soldier is a veteran already and gains none; a type with no veteran form stops at the full count (agent default, #1707 plan OQ1; enemy units count too, OQ2). An Israeli recruit that reaches the full count becomes a veteran_soldier with R-UNIT-004's promotion and its Why cause (no line for a single gain, OQ3), after the engagement, so it fought that engagement as a recruit. Every record change keeps the count (a refill too, R-TRAIN-004, OQ4), and a veteran stops accumulating service (OQ5). It draws nothing. The loader refuses either value when it is missing, Deferred or not a positive integer, with no fallback. The count is state, hashed as R-UNIT-002 says; a save replays its commands, so the save format is unchanged.

Corridors

Corridor work is step 5 too, after the sites (Phase 4 PR 2, #786; owner, 17 October 2026, #782: the plan's Q-COR-1 to Q-COR-7 and Q-RAIL-1). content/corridors.yaml holds each corridor's kind (road or rail), names, map route and segments; a segment's hexes are derived from the route's traced line by tools/corridor_hexes.py. The Money and turns per segment are labelled design values in rules.yaml mvp: (Q-COR-6). On screen a segment is its corridor's name and its number ("Burma Road, segment 2"), never a waypoint's name (Q-SENS-1). The MVP's corridors are the Burma Road (four segments), the 1949 Haifa–Tel Aviv rail link (one) and the Tel Aviv–Netanya road (four, on straight draft lines between its five sourced places; owner, 18 October 2026, #786, Q-COR-3 (b)). A segment point in a hex the grid counts as sea (the coast) counts in its nearest land neighbour, as a shore town does.

R-COR-001: PioneerWork(unit, build_road | build_rail, corridor:N) starts segment N (from 1) of a corridor whose kind the action builds (build_road a road, build_rail a railway), when the segment is neither being built nor built, with a halutz unit that has no other project (D-D6). Segments may be started in any order, one unit per segment, so several units can build one corridor at once (Q-COR-4); the unit works from where it stands (Q-COR-5). The Treasury pays the kind's road_segment_cost or rail_segment_cost at once and borrows the shortfall (R-MONEY-003). It is refused, with the state unchanged and the Attention refunded, for a busy unit, an unknown corridor or segment number, a kind the action does not build, or a segment being built or built; the legacy loop refuses it before charging Attention. Its Attention cost is 0 (D-A6).

R-COR-002: Each turn, after the sites, in ordinal corridor id order and then segment order, every segment being built gains progress x m of a turn of work (R-MIN-002), where progress is 1 or veteran_progress with a veteran (R-VET-002) and m is Transportation's multiplier, counting the turn of the order. Whole turns apply at once; the fraction is carried on the segment and hashed as 16 hexadecimal digits of its IEEE bit pattern only while non-zero, as in R-SITE-003. A segment whose progress reaches its kind's road_segment_turns or rail_segment_turns is built, on the turn's end date, and its unit is idle again. Corridors with a segment being built or built are written into the state hash only when there are some.

R-COR-003: A corridor is complete when every one of its segments is built, in any order. StopWork(unit) on a unit building a segment ends it as R-SITE-002 does for a site: the segment is unbuilt and free again, the unit is idle, and the progress and the Money paid are lost (D-A3). A built corridor needs no upkeep and yields nothing in the MVP (Q-COR-7).

Movement

Phase 5 PR 5 (#744). The numbers are design values this PR proposed (Q-MOVE-1, Q-MOVE-3; owner, 15 October 2026, #375: "Take all recommendations") and the owner accepted them on 19 October 2026 (https://github.com/eyalgolan/statecraft-israel/pull/1073#issuecomment-6259021508).

R-MOVE-001: Each unit type that moves has an integer movement_points a turn (unit_types.yaml; required, at least 1, on a military type; refused on a population type), and each land biome of map/hexes.json an integer cost to enter (rules.yaml mvp.move_cost, at least 1; the loader refuses a land biome with no cost and a cost for a water biome, so lake and sea are impassable). A path's cost is the sum of the costs of the hexes it enters. A unit's Moves in one turn share its points: a second Move has what the first left, and the next turn has them all again. The points left are the type's points less those the unit spent this turn (Movement.PointsLeft). A unit needs the full cost of the next hex, with one exception: a unit that has not moved this turn may enter one neighbouring land hex whatever it costs, and that move spends all its points (the owner, 23 October 2026, OQ-3: "No, strict like Civ VI", "with full points it can always move one hex"). A type with no movement points (the halutz types) cannot move. Real values: recruit 3 (owner, 19 October 2026, https://github.com/eyalgolan/statecraft-israel/pull/1130#issuecomment-6269103829), veteran soldier 3, artillery 2, armoured car 5; coastal plain, valley and lowland 1, hills and desert 2, mountains 3. Under exact identity gdd-0.3 (#1708), Movement.Allowance reads gdd03-balance.yaml military.move_infantry (2) or military.move_armoured (3), by type class for either army. The six land rows' movement pairs are infantry/armour: coastal plain, lowland and valley 1/1; hills and desert 2/2; mountains 2/3. These are existing Default catalogue values sourced to #1668 balance.md at 605277df. Reachability, path, route ETA, immediate execution, enemy movement, points-left, needs-orders and the UI denominator share this contract, including the full-points exception. Artillery keeps main's value, its type's movement_points (2), and pays the infantry column of those rows (the owner, 1 November 2026, Q12 on #1668: https://github.com/eyalgolan/statecraft-israel/pull/1668#issuecomment-6553884547); its fire spends the points it has left like any attack (R-COMBAT-001, R-COMBAT-007). No shared content bytes or older identity allowances change.

R-MOVE-002: Move(unit, q, r) takes the cheapest path, at command time (spec §2.3). Hexes leave the search in (cost, q, r) ordinal order, neighbours are tried in tools/hexgrid.py's NEIGHBOURS order, and a hex reached again at the same cost keeps the hex it was first entered from, so of two equal paths the one through the hex first in that order wins, and the path is the same every run. Movement.Reachable(c, s, unit, spent) and Movement.Path(c, s, unit, q, r, spent) search with the points left; without spent, with all of them. Movement.Reachable lists the hexes a Move can end on this turn with their costs, and Movement.Path the hexes a path enters; both come from the same search Move uses (spec §2.6). Movement.Route(c, s, unit, q, r, spent) is the cheapest path with no points limit, each hex numbered with the turn it is reached on: turn 1 spends the points left, later turns all of them, a step that does not fit what is left of a turn starts the next, and a step costing more than the full points is a turn of its own (R-MOVE-001). Under gdd-0.3, the same ordering and search consume R-MOVE-001's role costs and effective allowance; player actions still take effect at command time (owner ruling on #1708, 26 October 2026).

R-MOVE-003: A hex holds at most mvp.stack_limit units of one side (Q-MOVE-3; real value 3). The limit counts the moving unit's side only and applies where a move ends; a path may pass through a full hex. A hex never holds both sides' units: entering an enemy unit's hex is an attack (Phase 5 PR 8), so Move refuses it. Under gdd-0.3, the limit is the existing Default catalogue key military.stack_limit (3), including combat retreats.

R-MOVE-004: There is no zone of control (Q-MOVE-4): a unit passes next to an enemy unit at the terrain's cost, but no path enters or passes through an enemy unit's hex. Units may enter any land hex on the board; the 1948 control layer stays display-only. There is no transport (Q-MOVE-5). The same water, enemy-hex and no-zone-of-control restrictions apply under gdd-0.3; water catalogue rows must remain impassable.

R-MOVE-005: A player unit cannot enter the hex of an enemy-held locality that is not a curated target (settlement_targets.yaml, R-CTRL-001), and no path passes through one: in Phase 6 only Gaza stays closed (Q-CIV-1, owner, 15 October 2026: "No entry until Phase 6"; Q-TGT-1 (a)). The hex of an enemy-held target is open, so a player unit may move onto it, or through it, and capture it in play (R-CTRL-004; the owner's ruling on #793, 21 October 2026, that the targets "can be captured in play", https://github.com/eyalgolan/statecraft-israel/issues/793#issuecomment-6322289565; #1304). Raid's withdrawal (R-CTRL-010) and combat retreat (R-COMBAT-003) still never end on any enemy-held locality, target or not, and no enemy retreat ends on a player's locality (R-ENEMY-001). The localities are the settlements listed in rules.yaml mvp.enemy_held_localities (settlements_1948.yaml ids; the loader refuses an unknown one). The real list is the owner-accepted DRAFT (19 October 2026, https://github.com/eyalgolan/statecraft-israel/pull/1073#issuecomment-6259083180): acre, nazareth, lod, ramla, gaza, beersheba. It is unchecked: no person has yet checked each place's control on 15 May 1948; research task #1096 will check it. An annexed target Israel holds (R-CTRL-016) is no enemy-held locality: the player's withdrawal and retreat may end on it, and an enemy unit may enter it only as the hex its move is sent to, which takes it, and no enemy path passes through it (R-CTRL-006; the owner, 8 November 2026, answering the questions on #2955: "Can be taken" and "Destination only (Recommended)"). Under gdd-0.3 these guards remain; a kind: fort location grants neither locality capture eligibility nor peaceful entry onto hostile combatants (R-COMBAT-010).

R-MOVE-006: Move is refused with the state unchanged and a reason when the unit does not exist, is not the player's (a unit of a side other than Israel), cannot move or is already on the hex, when the hex is water or off the board, holds an enemy unit, is an enemy-held locality closed by R-MOVE-005 or is full of the unit's side, or when no path reaches it within the points the unit has left this turn. It costs no Attention (D-A6). A Move that goes through clears a fortify order (fortify, or fortifying the turn it is given) and the hex it names at once, so a unit that leaves its hex and comes back in the same turn is not fortified (Q-HOLD-1, "lost on move"); hold moves with the unit; a refused Move changes nothing. Under gdd-0.3 the same immediate/refusal/order-clearing contract uses the effective allowance and role terrain costs in R-MOVE-001. Moving never removes a permanent fort (R-COMBAT-010).

R-MOVE-007: Roads and rail are cheaper (D-G2; Phase 5 PR 6, #745). A step into a hex from a hex of the same built corridor segment (corridors.yaml, R-COR-002) costs the kind's mvp.road_move_cost or mvp.rail_move_cost (at least 1; the loader requires both in the GDD loop) when that is lower than the terrain's cost; otherwise it costs the terrain's. Only the corridors Phase 4 builds count, not the 1948 roads of map/routes_1948.yaml (Q-MOVE-2, owner, 15 October 2026, #375: option (a)). A segment not started or still being built costs the terrain, and a step from one segment onto another costs the terrain unless both hexes lie on one built segment. Real values: road 1, rail 1, design values Phase 5 PR 6 proposes; the owner may override them. A cost of 1 saves nothing on the coastal plain, valley or lowland, which already cost 1. Under gdd-0.3, the corridor costs come from Default catalogue keys military.move_road and military.move_rail (1 each); the same built-segment, both-endpoint and minimum-with-biome conditions apply.

R-MOVE-008: UnitOrder(unit, go_to, q, r) gives a standing go-to order and moves the unit at once as far as its points left reach along Movement.Route. It is refused with the state unchanged and a reason for an unknown or enemy unit, one that cannot move or is already there, water or off the board, a hex holding an enemy unit, a closed locality (R-MOVE-005), or a hex no route reaches. Otherwise it replaces the unit's order, including fortify. On arrival the order is cleared. A route may pass through a full friendly hex; at each turn's stop the unit uses the last hex with room. If the destination is full on arrival, the route stops and the order is cleared. Under gdd-0.3, R-MOVE-001 supplies allowance and role costs; no second allowance is applied at End turn.

R-MOVE-009: At the start of each player turn, before that turn's committed and waiting orders, every player's unit with a go-to order advances in ascending unit id along a freshly computed route. It spends movement points normally, including the full-points exception (R-MOVE-001). If no route reaches its target, or the target is full on arrival, it stops, clears the order and needs orders. An adjacent enemy does not stop it (Q-MOVE-4: no zone of control). If no hex it can reach this turn has room, it waits where it is and keeps the order. Under gdd-0.3, only an outstanding route continues at the turn boundary; ordinary moves and immediate attacks are not repeated, and the enemy follows the player's actions (owner ruling on #1708). Allowance and ETA follow R-MOVE-001.

R-MOVE-010: A unit needs orders when it is one of the player's military units, its type has movement points, it has neither moved nor attacked this turn, and it has no order (fortify, fortifying, hold and go_to are orders) (Movement.NeedsOrders, ascending id). End turn waits until no unit needs orders, apart from those the player skips this turn (the owner, 23 October 2026: "Block like Civ"; spec 2026-10-23 section 7). Skipping is not an order and changes nothing in the game.

Standing orders

Phase 5 PR 7 (#746). Q-HOLD-1, "(a) 1 turn, free, lost on move", and Q-HOLD-2, "(a) Own hex, no interception" (owner, 15 October 2026, #375). The two percents are design values this PR proposes from avlco's option 1 on #746; the owner may override them.

R-ORDER-001: UnitOrder(unit, order) gives one of the player's military units the standing order fortify or hold, replacing the one it had. A new order replaces a go-to too. The unit keeps its order, turn after turn, until the player gives another; moving clears fortify (Q-HOLD-1): in the resolver after combat (step 2), in id order, a unit no longer on the hex its fortify names (OrderTarget, "q,r") loses the order. The order is refused with the state unchanged and a reason for an unknown order or unit, a unit with no side (a halutz), a unit of another side, a hold for artillery (R-COMBAT-007), or the order the unit already has (fortify on the hex it fortifies). It costs attention_cost.unit_order (0, D-A6) and is refused in the legacy loop. A unit order names both hex coordinates or neither; one that names only one cannot be built (#2526), so no save can carry an order that live play refused.

R-ORDER-002: Fortify takes one turn to set and costs no Arms or Money (Q-HOLD-1). Given, it reads fortifying and names the unit's hex; in the same resolver step it becomes fortify, so the combat of the turn it is given sees no bonus. A fortified unit on its hex defends at mvp.fortify_defence_percent of its strength (Orders.DefencePercent; real value 125), and from the moment it leaves the hex at 100. Under gdd-0.3, the catalogue's Agreed military.fortify_turns must remain 1; settled fortify on its named hex adds the Agreed military.fortify_bonus (25) (O1, owner-approved 1 November 2026, #1668) to the other defence bonuses, once (R-COMBAT-002). fortifying adds nothing.

R-ORDER-003: Hold acts on the unit's own hex only and does not intercept or stop any enemy move, past or next to it (Q-HOLD-2). It names no hex and stays when the unit moves. A holding unit defends at mvp.hold_defence_percent of its strength (real value 100: hold keeps the position and adds nothing); a unit with no order defends at 100. Under gdd-0.3, Hold adds no defence bonus; it does not multiply terrain or permanent forts (R-COMBAT-002).

Combat

Phase 5 PR 8 (#747). Q-COMBAT-1 to Q-COMBAT-4 are the plan's recommendations (owner, 15 October 2026, #375: "Take all recommendations"); Q-CIV-2 is the owner's "Enemy units only, no effect" and Q-CIV-4 his "2+3". The numbers (rules.yaml mvp.arms_per_attack, mvp.combat_terrain_percent, mvp.combat_table), the loss order, the retreat order and the attack's standing are agent defaults asked on #747 (https://github.com/eyalgolan/statecraft-israel/issues/747#issuecomment-6297295556) and accepted by the owner on 20 October 2026, all six, as design values for the balance pass #813 (https://github.com/eyalgolan/statecraft-israel/issues/747#issuecomment-6299184133). The truce refusal (Q-WAR-3) is PR 11, prisoners PR 12, artillery PR 9 and the enemy's orders PR 10.

R-COMBAT-001: UnitOrder(unit, attack, q, r) makes one of the player's military units attack the hex (q, r) at once, alone, when the order runs (the owner, 23 October 2026: "Instant, like Civ" and OQ-1: "One unit per attack, Civ"). It replaces the unit's order (a fortify is lost), spends the rest of its movement points for the turn and leaves no attack order. The hex must be next to the unit (artillery: within its range, R-COMBAT-007) and hold an enemy military unit; enemy units on a hex with a locality and its residents may be attacked, and the residents are never a target (Q-CIV-2). The order is refused with the state unchanged and a reason for an unknown unit, a halutz, a unit of another side, no hex, a hex not next to it (artillery: beyond its range), no enemy unit, no movement points left or a stock below the Arms this attack needs. Fortify and hold name no hex: one that names a hex is refused. The order costs attention_cost.unit_order (0, D-A6). A save writes it as order <unit> attack <q> <r>; a scenario takes q and r. Under gdd-0.3, attacks and artillery fire use the same instant timing without an Arms affordability check or deduction; all other validity checks remain.

R-COMBAT-002: Resolver step 2 resolves the enemy's attacks in ascending attacker id: the lowest-id unit with an attack not yet in an engagement leads one, with every enemy unit attacking the same hex. A player's attack is fought when its order runs (R-COMBAT-001), as one engagement with that unit alone and one draw. Each unit's strength is first multiplied by its type's combat_percent (R-COMBAT-007, R-COMBAT-008). The attackers' strengths are summed; each defender's strength is multiplied by its order's defence percent (Orders.DefencePercent: fortify 125, R-ORDER-002) and by mvp.combat_terrain_percent for its hex's biome (real values: hills 125, mountains 150, every other land biome 100). The ratio is attack over defence in whole percent, rounded down, in integers only; the row of mvp.combat_table is the last whose min_ratio_percent it reaches (real rows: below 1:2, 1:2, 1:1, 2:1, 3:1). Each engagement draws exactly one RNG value, a face of the row (six in the real table), whatever its outcome; a step 2 with no engagement draws nothing (docs/determinism.md, "Combat draws"). Combat.Preview(c, s, unit, q, r) gives the ratio and row the UI shows (Q-UI-1; spec §2.6), counting that unit alone for a player's attack. Under exact identity gdd-0.3, each defender instead counts once at 100 + biome defence_percent + permanent fort bonus + applicable settled Fortify bonus percent, with bonuses read from the existing catalogue (#1668 military-combat.md:27, additive). Hold and unsettled/mismatched Fortify add nothing. Mountain + fort + settled Fortify is 225% of base defence (50 + 50 + 25); preview and resolution share this integer calculation. The older multiplicative formula remains unchanged. Under exact gdd-0.3, after the draw and the losses, retreats and surrenders, the engagement's survivors gain combat experience (R-UNIT-011), drawing nothing more.

R-COMBAT-003: The face gives the population units each side loses and whether the defenders retreat. Losses fall one unit at a time on the side's strongest unit in the engagement (a tie on the lower id; the attackers' losses spare their artillery, R-COMBAT-007); a unit at 0 is destroyed and leaves the state, and its id is never reused (NextId does not move). When the face says retreat, each defender left, in id order, moves one hex to the first neighbour in the fixed order (R-MOVE-002) that is land, holds no unit of the other side, has room for its side under mvp.stack_limit, and is not a locality closed to its side: for a player's unit an enemy-held locality (R-MOVE-005), for an enemy unit a player's locality (R-ENEMY-001, #2932). A defender with no such hex surrenders and leaves the state: the only way a unit is captured (impact plan §4.3 item 1; its people are counted as prisoners or as the player's soldiers held, R-POW-001, R-POW-005). The attackers never advance into the hex (Q-COMBAT-3). A player's unit's first loss stores the strength it had as its Full; only replacements (R-TRAIN-004) restore lost strength.

R-COMBAT-004: Each of the player's units pays mvp.arms_per_attack (real value 1; artillery pays its arms_per_fire instead, R-COMBAT-007) from the Arms stock when the attack is given; an attack the stock cannot pay for is refused. Enemy units pay nothing (Q-ENEMY-4: no supply). This price rule remains for gdd-0.1; under gdd-0.3, player attacks and artillery fire neither read nor spend these shot prices (#1706; owner ruling, 25 October 2026, on #1668). Arms stock, purchases, equipment, replacements and training refunds remain governed by their existing rules.

R-COMBAT-005: Combat changes only combatant units, the Arms and, through a surrender, the counts of prisoners and of the player's soldiers held (R-POW-001, R-POW-005), and, under gdd-0.3, the quiet-turn count of each front in the engagement (R-WAR-005). No city, locality, resident, building, job or other resource, and no score, changes because of a combat outcome (Q-CIV-2; impact plan §4.3: kills and enemy losses never score). Halutzim are never attackers or targets. Lost strength comes back only through replacements (R-TRAIN-004), which are training, not a combat outcome. Under gdd-0.3, attacks and artillery fire leave the Arms stock unchanged; the other combat effects in this rule remain. One exception, on a game with an initialized Historical ledger only: when a surrendering own formation has more serving records than its strength, the leftover records go home as R-POW-005 sets out (owner, 9 November 2026: "Sent home (Recommended)"), so they become residents, free civilians of the nearest city, and their issued items go back to the Arms stock.

R-COMBAT-006: Each engagement's Why line names the hex, the units on each side and the ratio, and states "the attackers lost N people and the defenders lost M people", the strength each side lost times people_per_unit, in the same words whichever side attacks (Q-CIV-4; impact plan §4.2). A destroyed unit gets "Unit N was destroyed", a retreat the hex it went to, a surrender the people it surrendered with. No running count of losses is kept; the war's Israeli total is PR 13's.

On a game with an initialized Historical ledger (the dark start, R-POP-004; #1706, #1725 PR 7; owner, 8 November 2026, #1725: https://github.com/eyalgolan/statecraft-israel/issues/1725#issuecomment-6726365722), Combat.Resolve and Combat.Outcomes run, and every engagement, step 2's and the player's instant attack alike, is fought as above on the units' strength, with the same draw, units and Why lines as the aggregate game, but for an own surrender's (below). Each population unit an own formation loses (R-COMBAT-003) also removes one of its serving soldier records, the active soldiers on its target (a starting force's units_1948/<row>, a trained formation's formation/<unit id> from R-TRAIN-003; "Yes, after #2964 (Recommended)"), from the ledger as a combat loss (R-POP-001): the lowest id first ("Lowest id first (Recommended)"), with the provenance combat/<turn date yyyy-MM-dd>/<q>,<r> of the engaged hex ("Date + battle hex (Recommended)"). Every loss is a death and there is no wounded status ("No, losses are deaths (Recommended)"). Combat experience stays on the formation (R-UNIT-011): no record's combat XP changes ("No, formation only (Recommended)"). A retreat moves the unit only: its records keep their formation's region-only location ("No, keep region (Recommended)"). A formation whose strength differs from its serving records still fights on its strength, and a record is removed only while one is left ("Go ahead on strength"). Enemy formations are not own population and their losses touch no record; a lost record's issued military item leaves with it, with no refund (R-COMBAT-005). An own surrender's losses are written before Prisoners.Take sees its records, so for a formation whose records matched its strength R-POW-005 finds the records left numbering its strength ("Prisoners #1710 first (Recommended)"); a formation whose records do not match is captured too, at most its strength of the records left, with no refusal (owner, 9 November 2026: "Go ahead on strength (Recommended)"; R-POW-005), and the records left beyond its strength go home as R-POW-005 sets out (owner, 9 November 2026: "Sent home (Recommended)"), with Send home's own Why lines after the surrender line ("Yes, reuse Send home's (Recommended)"). An own surrender's Why line states the people it surrendered with as the records captured times people_per_unit, not its strength times people_per_unit (owner, 9 November 2026: "Records captured (Recommended)"): the same key, surrendered_held, and the same one argument, a count of people; without a ledger it is the strength, as before. The losses draw nothing. A synthetic or Open ledger is still refused (R-POP-002). LedgerCombatTests pins this.

Phase 5 PR 9, part 1 (#748): artillery's combat role and the armoured car. Q-ARTY-1 and Q-ARTY-2 are the plan's recommendations (owner, 15 October 2026, #375: "Take all recommendations"). The numbers in unit_types.yaml are agent defaults proposed on #748 and accepted by the owner as design values on 21 October 2026, together with three of R-COMBAT-007's mechanics that the Q-ARTY-1 text does not settle: the attackers' losses sparing the guns, the hold refused as "never holds ground", and the 200% counting on defence too (owner, 21 October 2026, #748, https://github.com/eyalgolan/statecraft-israel/issues/748#issuecomment-6317339304). How either type is acquired (R-UNIT-008) is part 2, on top of Phase 5 PR 3's training.

R-COMBAT-007: Artillery (a type of class artillery, which needs a range and an arms_per_fire; real values 2 and 2) attacks alone from range. Its UnitOrder(unit, attack, q, r) names an enemy hex at a hex distance of 1 to its range; a farther hex is refused with "That hex is beyond the unit's range". The player's artillery attacks at once like any unit, pays its type's arms_per_fire in place of mvp.arms_per_attack, and its losses spare it, so it loses nothing. Enemy artillery still joins an engagement in step 2 after the other attackers of its side and only when at least one of them joins; otherwise it holds its fire, with a line, pays nothing and keeps its order. Its strength counts at its combat_percent (real value 200). Artillery that is attacked defends and takes losses like any unit. It never holds ground: a hold order for it is refused with "Artillery does not hold ground" (fortify is allowed). As in any attack, residents are never a target (Q-CIV-2, R-COMBAT-005). Under gdd-0.3, artillery still fires alone at once from the existing range and keeps the same outcomes, but has no Arms affordability check or deduction; arms_per_fire is not required or read in that identity. Range and combat factors retain their existing validation; the catalogue's combat.artillery_range (2) is Agreed (O5, owner-approved 1 November 2026, #1668), and its combat.artillery_factor is Agreed at 2, the same 200% as combat_percent (C1, owner, 10 November 2026, #2918: "200% wins (Recommended)"). The remaining combat redesign is held.

R-COMBAT-008: A military type may carry a combat_percent (at least 1; none is 100): its units' strength counts at that percent in the ratio, attacking and defending, in integers (R-COMBAT-002). The armoured car's is 150 (real value), with its 5 movement points (R-MOVE-001) and no other rule (Q-ARTY-2). Under gdd-0.3, armour's movement allowance is the catalogue value 3 (R-MOVE-001); combat factors and other mechanics remain unchanged.

Phase 5 PR 11, slice 2 (#750): the truce cards bind the player's attack orders (Q-WAR-3). The windows and the two rulings below are the owner's of 21 and 22 October 2026 (https://github.com/eyalgolan/statecraft-israel/issues/750#issuecomment-6327133246). The card-to-window binding by ChosenOptions is an agent default of this PR. R-ENEMY-006's date-driven enemy guard remains separate from the player's card-driven truce state.

R-COMBAT-009: From accepting a truce card until its window ends, the player's attack orders are refused with "A truce is in force": from first_truce's accept on the turns that start before 8 July 1948, and from second_truce's comply on the turns that start before 15 October 1948, on the sourced dates of events/first_truce.yaml, events/ten_days.yaml, events/second_truce.yaml and events/operation_yoav.yaml (R-ENEMY-006's windows). A refused or defied card binds nothing: only the accepting option starts the binding. A complying player attacks again from the 15 October 1948 turn (owner ruling 1). The rule reads its turn by the turn's first day (the date the HUD shows) and R-ENEMY-006 by the turn's middle day; on monthly turns the two agree: the turns of 15 June to 15 September 1948 are bound, 15 October is free. Movement and every other order stand. The enemy side's own guard is untouched. The truce's refusal takes nothing: the order names its hex before the truce is checked, the state and the Attention are preserved, and no Arms are paid.

R-COMBAT-010: Under exact identity gdd-0.3 only, an effective map-location hex with at least one existing kind: fort record grants either defending army the Default catalogue military.fort_bonus (50 percent). Existing records are latrun_fort and iraq_suwaydan_fort. Q13 says: "Every kind: fort record is a permanent fort", and adding one needs a sourced location row. Use effective placement after map overrides, with a legal land hex required. Duplicate markers grant one bonus. A fort is immutable content, independent of its occupant: vacancy, departure, re-entry, retreat, destruction and a change of defending army do not remove it. Attacking from a fort gives no attack bonus. No managed-city, capture-target, control, defeat, save schema or mutable-state property is created. Bonuses combine as R-COMBAT-002 specifies; older identities ignore permanent forts. Source: #1668 balance.md:141 at 605277df and the existing locations_1948.yaml records, reused without new historical claims.

The enemy

Phase 5 PR 10 (#749): part 1 is #1263 and part 2 is #2087. Q-ENEMY-1 to Q-ENEMY-4 are the plan's recommendations (owner, 15 October 2026, #375: "Take all recommendations"). The march's turns (through the 15 June 1948 turn, until the first truce ends on 8 July) are the owner's (owner, 21 October 2026, #749); so is the truce guard (R-ENEMY-006, same ruling). The scripted anchor's choice, the goal order, mvp.enemy_attack_ratio_percent and the truce guard's turn-middle reading are agent defaults proposed on #749 and #1263. The operation cards' orders (Q-ENEMY-5, the plan's R-ENEMY-005) are part 2, on main since #2087 (merged 3 November 2026): its ruleset, its byte pins and its orders are the owner's (31 October 2026, #749), and since #2621 R-ENEMY-007 applies them under both GDD identities.

R-ENEMY-001: Resolver step 1, before combat, gives every enemy military unit its orders, in ascending unit id, from the state alone: two equal states get the same orders. They go through the player's own checks with the sides swapped (Movement.Move, Combat.Give, Orders.Give), and a refused order changes nothing. So an enemy unit may not enter a hex holding a player unit, water or a full hex (R-MOVE-003), and, as R-MOVE-005's mirror, no locality's hex that is not on mvp.enemy_held_localities, except the player's locality or annexed target a move is sent to, which it then takes (R-CTRL-006; the owner, 8 November 2026, answering the questions on #2955: "Can be taken"). No enemy path passes through a player's locality or an annexed target Israel holds (the owner, 8 November 2026, answering the questions on #2955: "Destination only (Recommended)"), as a card's go-to goes round every place it is not sent to (R-ENEMY-005; the owner's answer 1 on #2087, 31 October 2026), and no enemy retreat ends on one. An arab_town off the list (Jaffa) and a hex it shares stay closed. Its attacks pay no Arms (Q-ENEMY-4: no enemy supply). Step 1 draws nothing from the RNG (Q-ENEMY-2).

R-ENEMY-002: In Historical mode each starting Arab force (R-UNIT-006) has a script from map/units_1948.yaml (Q-ENEMY-1; research_notes/GDD phase 5/arab-forces-1948.md section 10). Its 15 May 1948 position is the start. Its script is every anchor of the row from 11 June 1948 on that cites at least two sources other than Wikipedia and Wikidata, each on its hex (map/unit_anchor_hexes.yaml), in the file's order. On each turn its target is the first of them whose window has not ended when the turn starts (its to; owner, 31 October 2026, #1270 ruling a: "First listed (as today)"): to 8 July 1948 Isdud for Egypt (the forward base, listed before Majdal), Latrun for the Arab Legion, the bridgehead at Mishmar HaYarden for Syria and Jenin for Iraq; the ALA has none. After 8 July only Syria has one, the same bridgehead to the Syrian armistice of 20 July 1949 (#1270, anchor S1; the other forces' anchors past the first truce are #1270's later slices). The unit moves along its cheapest path to the target (R-MOVE-002) as far as its movement points reach. A unit on its target, with no path to it or whose move is refused follows R-ENEMY-003 that turn, and so does every unit from the turn its last window has ended. The script never places a unit anywhere but its own path to a sourced anchor.

R-ENEMY-003: An enemy unit off its script attacks the first neighbouring hex, in the fixed order (R-MOVE-002), holding a player military unit where the enemy attack preview (Combat.EnemyPreview, its own strength plus its side's units that already carry a standing attack on that hex and can still reach it) reaches mvp.enemy_attack_ratio_percent (real value 200, the 2:1 row; a labelled design value). Under exact gdd-0.3 the threshold is the catalogue's combat.enemy_attack_ratio instead, 1.5:1 (150): owner, 1 November 2026, O4 on #1668, "1.5:1 in gdd-0.3 (Recommended)" (#2114). A standing attack on that hex is kept. Otherwise it fortifies its hex (R-ORDER-002; a fortify it has is kept). Step 2 resolves its attack in the same turn. Inside a truce it only fortifies (R-ENEMY-006).

R-ENEMY-004: No enemy unit arrives or grows in Phase 5 (Q-ENEMY-3): no turn adds an enemy unit, until a dated per-army figure is chosen under CONTENT_POLICY section 1.

R-ENEMY-005: An option of an event may give go-to orders to units (unit_orders, Q-ENEMY-5; R-ENEMY-008 adds attack orders): each names a map/units_1948.yaml row and a hex that must be one of that row's anchor hexes or order target hexes in map/unit_anchor_hexes.yaml, or the loader refuses it, naming the file and the field. Once R-ENEMY-007 permits the operation to start, each order in list order goes to the row's starting unit (R-UNIT-006: the unit with id i is the i-th starting force) through the player's own go-to (R-MOVE-009), with the movement points the turn's orders have already spent; a refused order changes nothing, as the player's does, and the turn's refusals list it as the player's refused orders are listed (Orders.RefusedCardOrder, shown in English and Hebrew). A card's order never takes the last Israeli military unit off an occupied place, occupied or awaiting its end (R-CTRL-013; owner, 2 November 2026, #2087: "Keep a garrison in Nazareth (Recommended)"): it is not given to that unit, and on any later turn of its march a card's order whose unit is now the last one there is taken away at the route step, the unit staying where it is; either way the turn's refusals list it with Orders.RefusedLastGarrison, as the other refused orders are listed. A second Israeli unit on the place lets the order go ahead, and the player's own go-to is never held back. Each unit is checked on its own, when its turn comes: in the option's list order when the card gives its orders, and in ascending unit id at the route step (R-MOVE-009). So of two Israeli units on one occupied place that both march on a card's orders, the one checked first goes, and the other, then the last one there, stays: at the route step the lower id goes and the higher id stays (EnemyAITests.OfTwoUnitsOnACardsMarchTheHigherIdStays). A held order is taken away, not kept for later: a unit that arrives on the place afterwards does not restart it. So when Horev is launched with Golani alone on an occupied Nazareth, Golani stays and the occupation goes on. Only sourced anchors receive orders (owner, 31 October 2026, #749: "Sourced anchors only"), and every military operation gives some (owner, 1 November 2026, #2087: "military operations must have unit orders"). A place an operation sent a unit to that is not one of its anchors is a row's order_targets entry (owner, 2 November 2026, #2087, option 1): sourced and dated as an anchor, read only by the cards' orders and tools/unit_anchor_hexes.py, never drawn and never a start; the loader refuses one with no sources or no from/to date. Kiryati's Ramle, Givati's Huleiqat, the Negev Brigade's Beersheba, Carmeli's Manara, the 7th Brigade's al-Malikiyya and Harel's Latrun police fort are order targets: ten_days's historical option sends Kiryati to Ramle (Operation Dani; the owner's "Kiryati to Ramle", which replaced Yiftach's order to Lydda, as Yiftach has no unit on the board) and Carmeli to face the Syrian bridgehead at Mishmar HaYarden (Operation Brosh); operation_yoav's sends the Negev Brigade to Beersheba and Givati to Huleiqat; operation_hiram's sends Carmeli to Manara and the 7th Brigade to al-Malikiyya, which is refused while the ALA's unit holds that hex, as in the historical run; operation_horev's sends Golani to the Julis camp; and latrun_proposal's press_offensive sends Harel toward the Latrun police fort, the objective of Ben-Gurion's proposal of 26 September 1948 (owner, same day), refused while an enemy unit holds the fort. Kiryati's Ramle target is Ramla's capture target (-3, 24), not the (-4, 24) Ramla shares with Rehovot on the 4.4 km grid (owner, 2 November 2026, #2087: "Send it to Ramla's target"; tools/unit_anchor_hexes.py). Every other alternative option gives none until sources are found. The orders apply under both gdd-0.1 and gdd-0.3; the legacy loop has no units and ignores them. Orders expire with the operation (the owner's answer 1 on #2087, 31 October 2026): a card's order ends on the to date of the anchor or order target it names, its window in map/units_1948.yaml (Carmeli's Brosh anchor to 19 July 1948, Kiryati's Ramle target to 16 August 1948, one turn past Operation Dani's end on 19 July, so the order takes Ramla on the 15 August turn and not Lod (owner, 2 November 2026, #2087: "Longer order: takes Ramla (Recommended)"), Givati's Huleiqat target to 1 November 1948, Harel's Latrun target to 22 December 1948, as its Sha'ar HaGai window, Golani's Julis anchor and the Negev Brigade's Beersheba target to 24 February 1949, the Egypt armistice, and Carmeli's Manara and the 7th Brigade's al-Malikiyya targets to 23 March 1949, the Lebanon armistice), and from the first turn on or after that day the unit stops where it stands (card_order_ended), arrived or not; the loader refuses an order to an anchor with no to. The anchor's dates, not the card's: the card's window is when the decision is open. A card's march goes round every enemy-held locality (R-MOVE-005, open target or not) but the hex it was sent to, so it never ends on, or captures, a place it was not sent to; a route that cannot go round is refused. The player's own go-to, which may end on an open target and capture it, is unchanged, and replaces a card's order. The four cards keep their review status (owner, same ruling: "Re-pin; you sign off").

R-ENEMY-006: No enemy unit attacks inside a truce window (owner, 21 October 2026, #749): the first truce from 11 June to 8 July 1948 and the second from 18 July to 15 October 1948, half-open, the dates of events/first_truce.yaml, events/second_truce.yaml, events/ten_days.yaml and events/operation_yoav.yaml (the same windows as map/control_hexes_1948_49.yaml's truce periods). A turn is in a truce when its middle day is (agent default): a monthly turn spans a truce's edge, and the middle puts it on the side that holds most of its days. So the turns of 15 June to 15 September 1948 are truce turns, and those of 15 May and 15 October 1948 are not. On a truce turn every enemy unit's standing attack ends first, whatever the unit then does (owner, #1263): it may still move on its script (R-ENEMY-002), and a unit off its script fortifies (R-ENEMY-003's else branch). The guard remains date-driven in every mode, whatever the truce cards' answers; #750 separately made accepted truce cards bind the player's attacks (R-COMBAT-009).

R-ENEMY-007: The owner, 4 November 2026 (#2543): "A military operation can only get underway if you have the required units in place (you cant fight in latrun in you dont have military units there)." For #2543 the owner chose the options labelled "In the same region", "Both rulesets (Recommended)" and, later, "Player must move them"; for a delegated card the owner's words, as recorded on #2543 on 4 November 2026, are "Postpone and explain to the user which units need to move where. Postpone until they are in the right area and if the operation is still relevant". Under the agent-recorded scope, every option with unit_orders has a non-empty requires_unit_regions list. Its required units are exactly the distinct starting-force rows in its unit_orders; each must still be on the board and stand on a hex in any one listed region when the option is decided. No starting position receives special credit and the sim does not pre-position a unit. The five requirements recorded in #2543 are: ten_days.central_front, Kiryati and Carmeli in central_plain, jerusalem_hills or galilee; operation_yoav.stop_at_ceasefire, Negev and Givati in southern_coast or negev; operation_hiram.stop_at_ceasefire, Carmeli and 7th Brigade in galilee; operation_horev.launch, Golani in southern_coast or negev; latrun_proposal.press_offensive, Harel in central_plain or jerusalem_hills. The owner's rulings of 7 November 2026 ("Add real orders", "Same units (Recommended)", "Same units as Central (Recommended)", "Keep the same targets (Recommended)", "Stop at al-Malikiyya (Recommended)") add three more, each with its units' sourced targets: operation_yoav.push_to_hebron, Givati to Bayt Jibrin (from 27 October 1948, the later of the sources' dates, "Second read, then later date (Recommended)") and Negev to Beersheba, in southern_coast or negev; operation_hiram.advance_into_lebanon, Carmeli to Manara and 7th Brigade to al-Malikiyya, in galilee, the crossing into Lebanon told by the card only; ten_days.southern_front, Kiryati and Carmeli, in central_front's regions and to its targets, Ramle and the north, not a southern target. A direct unmet choice is refused before any state changes. If delegation draws an unmet operation option, the card remains pending, the turn report (its operations row) and the Why screen say which units must move to which regions, and the draw is rolled back. Since the owner's rulings of 7 November 2026 ("All five cards", "Wait, with a limit (Recommended)", "6 turns (Recommended)", "Keep the option (Recommended)"), the postponed card keeps that option and later delegations draw nothing (the state's postponed{} records the option and the turn and date of the first postponement); it is shown again on 6 turns (mvp.operation_wait_turns) and lapses at the end of the 6th, without a chosen option, effects or orders. A card never postponed lapses when its window closes, as before. A postponed card whose historical option needs no units takes that option at its next delegation, and its wait ends (Latrun: "Takes the vote, wait ends (Recommended)"). A postponed card lapses at once when a unit its option needs is no longer on the board ("Lapse at once (Recommended)"; a dated row is waited for until its day), and any pending operation lapses when its front's armistice is signed ("Signing wins (Recommended)"; armistice_front: Yoav and Horev egypt, Hiram lebanon, "Keep Lebanon (Recommended)", Ten Days and Latrun jordan, "Jordan signing (Recommended)"). While that front's armistice card is pending, the operation cannot be chosen and its delegation postpones it ("Decide armistice first (Recommended)"). Every lapse costs mvp.operation_lapse_security (-0.01) in Security, Horev's Hold ("All five -0.01 (Recommended)"). An operation that waited and launches while a truce the player accepted is in force costs mvp.operation_truce_break_standing (-0.10) in International Support, unless its own effects already cost it, as Latrun's offensive does ("Same as Latrun, -0.10 (Recommended)", the owner's follow-up on a late Ten Days). Its card orders end later by the time it waited ("Shift by the delay (Recommended)"; the coordinator's reading, from the first postponement to the launch). Unlike an ordinary card under R-EVT-001, an operation card never first appears once its window has closed, and the wait for a soft slot does not keep it past its window. A card with follows (Rhodes Talks, Sinai Withdrawal, Rafah Force Pressure) fires on the turn after its requires are met even when its own window has closed ("Follow the launch (Recommended)", "Shift with Sinai (Recommended)"), and a card with armistice_front never first fires after that front's armistice is signed, so these three drop if Egypt signs before they are shown ("They drop (Recommended)"). A card shown after its own window has closed (a late follow-up, or an operation shown again while it waits) shows its late_text, and in Hebrew its late_text_he: Rhodes Talks, Sinai Withdrawal, Rafah Force Pressure, Operation Yoav and Operation Hiram have one ("Dateless variants (Recommended)"; the texts, "All approved (Recommended)", 7 November 2026). Operation Horev's text now says Golani must reach the south first ("Reword it (Recommended)"). This applies under both GDD identities. All requirement and missing-row iteration is ordinal; immutable region geography is loaded under both identities, while regional population/control state remains gdd-0.3 only. Under gdd-0.1 the loader reads regions.yaml's geography for this rule but no field host: the host rows (R-POP-004, #1717) and their checks stay exact gdd-0.3, as the opening field hosts' section says.

R-ENEMY-008: The owner, 4 November 2026 (#2544): "In addition, deciding on a military operation makes the units do what the operation describes (shown to the user all the moves, attacks etc)." An option's unit_orders row may also be an attack: { unit, order: attack, q, r } (order is go_to when left out). Its hex, like a go-to's, must be one of the row's anchor or order target hexes, so a card attacks only where a source puts its unit's fight (the maintainer's decision on #2544: only actions a card's cited sources state are implemented, and nothing is inferred from its prose); it needs no to date, as it is fought when the card is decided (the owner, 8 November 2026, answering the maintainer while #2956 was in review: "When card decided (Recommended)"); a target out of reach then is refused, and the unit does not wait to attack on arrival. A row may appear once as a go-to and once as an attack in one option, never twice as either. When the option is chosen or delegated and R-ENEMY-007 permits it, the rows run in list order: a go-to as R-ENEMY-005 gives it, and an attack through the player's own attack (Combat.Strike, R-COMBAT-001): the same checks and refusals (the hex in reach, an enemy unit on it, movement points left after the turn's earlier orders, no accepted truce in force under R-COMBAT-009, and the Arms under gdd-0.1), one engagement and one draw, and the attacker's points spent; the last-garrison rule is a go-to's only, as an attacker never leaves its hex. As the player's attack does, it ends the attacker's order, so list order matters: a row that goes to a hex and then attacks marches as far as its points allow, attacks, and its march ends there; a row that attacks and then goes to a hex keeps the go-to and marches from the next turn. A refused attack changes nothing and draws nothing, and the turn's refusals list it as a card's refused go-to is listed (Orders.RefusedCardOrder). When an action is refused, the owner chose the option labelled "Continue the rest (Recommended)": the later actions still run. The owner chose the option labelled "Both rulesets (Recommended)": this applies under gdd-0.1 and gdd-0.3; the legacy loop has no units and ignores it. A delegated card's attack draws from the same stream as the delegation's own draws, after them, and the draw is kept. The sim lists every action an operation takes, in the order it ran, with its card, option, row, unit, order, hex and outcome (OperationAction: carried out, or refused with the player's own reason), on the turn's StepResult and PreviewResult, for the map and the turn report to show (#2544's later pull requests). No card has an attack row yet: each waits for a research note citing the attack (#2135).

R-ENEMY-009: The owner, 4 November 2026 (#2544): "Also if a user tries to cancel orders of a unit that is in an operation, a message would tell him that this will cancel the entire operation if he choses to do this". A unit is in an operation while it carries out that card's go-to (R-ENEMY-005): the decided card whose option sends the unit's row to the hex it is going to (EventSystem.OperationOf; the loader refuses two cards that send one row to one hex, so the card is known). An order the sim accepts for such a unit that takes it off that go-to (a go-to of the player's own, an attack, fortify, hold or send home; a single move keeps the go-to) cancels the operation. When the player cancels a unit's orders in an operation, the owner chose the option labelled "Only remaining orders (Recommended)": only the orders still outstanding are cancelled, and the prompt says "remaining orders". Completed moves and combat stay. Every other unit still carrying the operation's go-to stops where it stands, with the line of an ended card order (card_order_ended; the owner, 8 November 2026, answering the maintainer while #2956 was in review: "Reuse existing line (Recommended)"), and the sim lists each of the operation's units, the ordered unit too, as withdrawn (OperationAction). A refused order cancels nothing. Another card's order that replaces a unit's card order moves the unit to that card's operation and cancels nothing (the owner, 8 November 2026, answering the maintainer while #2956 was in review: "1 + it can effect the success of the first op as it has less forces"). The first operation goes on without the unit: its remaining go-tos bring one unit fewer to their hexes. Its attacks were all fought when its card was decided, each by its unit alone (R-COMBAT-001, OQ-1), so none is left to weaken; the sim keeps no success or failure of an operation as a whole. Asked whether that is enough, the owner chose the option labelled "Units' fights are enough (Recommended)" (8 November 2026, recorded on #2956): a smaller operation is weaker only through its units' own fights and moves, with no new mechanic. The confirmation the player sees before such an order is #2544's user-interface pull request; the sim's cancel draws nothing and applies under both GDD identities.

The war state

Phase 5 PR 11, first slice (#750; Q-WAR-1, Q-WAR-2 and Q-WAR-4, owner, 15 October 2026, #375: "Take all recommendations"; D-C4): the security rung stops being a player lever in the GDD loop and stays the value the war events set. The player's attacks in a truce (Q-WAR-3) are a later slice of #750.

R-WAR-002: In the GDD loop the sim refuses SetMobilization and SecurityOrder with "That order is not available in this game", the state and the Attention unchanged (the convention of R-CITY-001's legacy refusal). The call-up follows the rung instead: when the rung changes (an event's set_rung), the call-up becomes that rung's rung_min_mobilization, raised as before and now also lowered, with the cause "Reserves sent home: call-up lowered to N%, the minimum for security level R". That is what HistoricalAI's call-up order did a turn later in the legacy loop (the army demobilised after the 1949 armistices), so the debt calibration holds (EconomyCalibrationTests, the Michaely IL 98.4m 1950-51 anchor within 15%; the owner's condition on #1148). HistoricalAI sends no call-up order in the GDD loop, and the Defence screen hides its escalate, de-escalate and call-up controls there. The legacy loop accepts both orders and keeps the call-up where the player put it.

R-WAR-003 (first slice): the GDD loop retires the 52-week UN ceasefire clock (Q-WAR-2, option (b)): it never starts (CeasefireWeeksLeft stays 0), never counts down and never ends the war, and the Defence screen shows no ceasefire line there. An event's set_rung still moves the rung (Q-WAR-4). ceasefire_weeks stays in the rules for the legacy loop, which keeps the clock.

R-WAR-004: Military actions move no parameter in the GDD loop (Q-CIV-3, "No effect in Phase 5", owner, 15 October 2026, #375): attacks, attacks refused in a truce and moves (the sim has no frontier, Q-MOVE-4, so crossing one is a move), with their losses, leave International Support (intl_standing) and Social cohesion (morale) exactly where a turn without them does. R-POW-002's prisoner Food-shortfall cohesion cost is the stated exemption.

R-WAR-005: Under exact gdd-0.3 (gdd-0.1 in #752), each of the four fronts, Egypt, Jordan (with the Iraqi sector), Lebanon and Syria, keeps a count of consecutive quiet turns from the scenario start, so turns before 16 November 1948 count (Q48, owner, 1 November 2026). A quiet turn has no engagement with a unit of that front, whoever attacked: the player's instant attack and step 2's attacks alike (Combat.Engage). A mixed-front engagement breaks the count of every front in it; a refused order engages nothing and breaks nothing; the ALA and local forces belong to no front. At the turn's end, after the war loss, each front not engaged gains 1 and each engaged front goes to 0. An early armistice offer is eligible for a front not yet signed from the 15 December 1948 turn (owner, 27 October 2026, #1709: "The first offer can come on the 15 December 1948 turn, not earlier") once the count reaches military.armistice_quiet_turns (2, content/gdd03-balance.yaml). The offer itself is R-WAR-008. The counts and the fronts engaged in the open turn are written into the state hash only when present, so no other ruleset's hash changes. Draws nothing.

R-WAR-006: Under exact gdd-0.3 (gdd-0.1 in #752), refusing a front's armistice (hold_out on armistice_<front>) sets that front's count to 0, so two full quiet turns must follow before an early offer is eligible again (Q49, owner, 1 November 2026; Q-OBJ-2). The refusal's own turn does not count (Q49: "Does the refusal turn count toward the next quiet period? No."): the card is answered before its turn resolves, so the refusal also marks its front as engaged for that turn, which ends at 0; a refusal on turn T with T+1 and T+2 quiet makes the front eligible on T+3. Refusing the armistice card is the refusal that resets the count, as Armistice.Resolved already does (owner, 9 November 2026, #1709: "Yes, keep (Recommended)"); this closes #1709's open question 4.

R-WAR-007: Under exact gdd-0.3 (gdd-0.1 in #752), signing a front's armistice (sign on armistice_<front>) removes every unit of that front from the board, Iraqi units with Jordan's, and no other (Q-OBJ-4, owner, 15 October 2026). Prisoners and the player's soldiers held stay as they are until the exchange card (R-POW-004), and Send home follows R-UNIT-005 unchanged: a unit whose nearest front-bearing enemy is on a signed front may go home. Because the signed front's units leave, a unit that also borders another front's unit now counts as on that other front and waits for its armistice. The armistice lines stay display-only.

R-WAR-008: Under exact gdd-0.3 (gdd-0.1 in #752), each front's armistice card (armistice_<front>, front_quiet: <front>) also comes while that front may have an early offer (R-WAR-005, R-WAR-006): before its from, and again after a refusal (#2478). It reads only its own front's count, so another front's quiet turns never bring it. On that path it is a soft event that takes the first soft slot (R-EVT-002). The sourced from stays the card's ordinary path, a hard anchor as before, so a front that never turns quiet gets its card on its date. A refused card, early or on its date, comes back only once two full quiet turns have followed the refusal, and its from does not bring it back again; a signed front gets no card again. Left unanswered, the card is still a hard anchor for R-EVT-003, so in Historical mode it takes its historical option (sign) on either path, as early as the 15 December 1948 turn (owner, 4 November 2026, #2508, option A: "keep early signing", https://github.com/eyalgolan/statecraft-israel/pull/2508#issuecomment-6630316010). Shown before its from (the early path), a card with an early_text shows it, in Hebrew its early_text_he: the talks came sooner than in history. The card is then marked as an alternative course by the event.alt_early banner (owner, 9 November 2026, #1709: "Separate early text (Recommended)", "Banner (Recommended)"). A card shown on or after its from, on either path, shows its text, so a refused card that comes back then shows its ordinary text and no saved state records the path ("Ordinary text (Recommended)"). Its dates, options and effects are the same on both paths. Draws nothing.

Prisoners

Phase 5 PR 12 (#751). The impact plan's §4.3 as written (Q-POW-1), points per unit returned (Q-POW-2), a sourced event card per front (Q-POW-3), the player's soldiers counted with a cost (Q-POW-4) and a Food cost whose shortfall costs cohesion (Q-POW-5): the owner's picks of 15 October 2026 (#375). The four amounts (rules.yaml mvp.food_per_prisoner_unit, mvp.prisoner_shortfall_cohesion, mvp.points_per_prisoner_unit_returned, mvp.held_soldier_cost_per_turn) are agent defaults posted on #751 (https://github.com/eyalgolan/statecraft-israel/issues/751#issuecomment-6318787961), accepted by the owner as design values (owner, 21 October 2026, #751: https://github.com/eyalgolan/statecraft-israel/issues/751#issuecomment-6321366548). The same ruling says ALA and local-force prisoners belong to no front, so they stay held, and fed, to the game's end. It does not name the front of each side or the captor of the player's soldiers; those stay agent defaults. The four sourced exchange cards are not in content/ yet: they wait for a person's source check (CONTENT_POLICY §2) and the owner's sensitivity review.

R-POW-001: Prisoners come only from a surrender in combat (R-COMBAT-003). The surrendered enemy unit's strength, in population units, is added to GameState.PrisonersHeld under its side; residents never become prisoners (R-CTRL-003). No command and no event effect adds to or takes from the count: no effect target names it. The surrender's Why line ends "They are held as prisoners", and nothing on screen celebrates a capture (impact plan §4.3 item 4). On the dark start's Historical ledger (#1710, #1725 PR 7) enemy prisoners stay this count per side, with no ledger record (owner, 8 November 2026: "No, keep count (Recommended)").

R-POW-002: In step 7, after the nation eats and the import (R-FOOD-002), the prisoners eat mvp.food_per_prisoner_unit (real value 1, one unit's ration) for each population unit held, as far as the stock goes. Each unit of their Food the stock cannot give lowers Social cohesion by mvp.prisoner_shortfall_cohesion (real value 0.01), the one cohesion cost of a military matter in Phase 5 (Q-CIV-3); a shortfall never reduces the prisoners (Q-POW-5).

R-POW-003: Prisoners are never labour, Money, Arms, housing occupants, jobs or a resource any command or event can spend: a state with prisoners and one without end every command's turn with the same Money, Arms, cities and units, and Food that differs by the upkeep alone.

R-POW-004: An event with prisoners: <front> (egypt, jordan, lebanon or syria) is that front's exchange card. It fires only while the front holds prisoners either way, a stored 0 holding nobody (the Jordan card also returns the Iraqi prisoners: the Jordan agreement took over the Iraqi sector, Art. VI, Q-OBJ-1; the ALA and the local forces belong to no front, so no card returns theirs). It has exactly one option and resolves without an RNG draw, like an objective's card. Resolving it moves its sides' prisoners from PrisonersHeld to PrisonersReturned and adds mvp.points_per_prisoner_unit_returned (real value 10) to the score for each population unit returned (Q-POW-2). Nothing scores at the capture (impact plan §4.3 item 3). It holds unchanged under gdd-0.3, played with a synthetic exchange card: a returned unit scores 10 (owner, 1 November 2026, O7: "10 per unit, main (Recommended)"; PrisonersGdd03Tests.AReturnedUnitScoresTenPoints); the card fires once, so prisoners taken after it has resolved stay held to the end of the game with no points (27M-19; PrisonersGdd03Tests.PrisonersTakenAfterTheCardStayHeldWithNoPoints); and a loaded save pays no return twice (PrisonersGdd03Tests.ASaveAndReloadCannotPayTwice).

R-POW-005: A surrendered unit of the player's is added to GameState.SoldiersHeld under the side of the engagement's lowest-id attacker. In step 7, after the prisoners eat, the Treasury pays mvp.held_soldier_cost_per_turn (real value IL 0.005m) for each population unit held, borrowing what it lacks (R-MONEY-003; Q-POW-4). They come home when that side's front's exchange card resolves, with no score effect. On the dark start's Historical ledger (#1710, #1725 PR 7) the surrender also makes the formation's serving soldier records Captured, in one validated ledger commit after every check. Their captor is that same side: every enemy side with military units (a unit_types.yaml side other than israel and "") is bound as a captor when the Historical ledger is initialized (owner, 8 November 2026: "Winning side, set at start (Recommended)"). Their Home is the formation's region, region only ("Formation's region (Recommended)"). Each keeps its id, history and issued military item, unchanged while held (docs/GDD.md Q10; "Stays on record (Recommended)"). GameState.SoldiersHeld keeps its count beside the records, and per captor side it must equal the Captured records with that captor: Take and Upkeep check this before any write ("Keep count + check (Recommended)"). The held soldiers keep this Money cost ("Yes, as today (Recommended)"; docs/gdd/military-combat.md calls the exemption non-operative). The records captured are the active soldiers on the formation's target, found as battles find them (Training.FormationTarget: an opening formation's units_1948/<row>, a trained formation's formation/<unit id> from R-TRAIN-003), so a trained formation's soldiers become captives too (owner, 9 November 2026: "Yes, in #1706 (Recommended)"). A formation whose serving records do not number its strength is not refused, and the turn goes on (owner, 9 November 2026: "Go ahead on strength (Recommended)"): with more records than its strength, its strength of them are captured, the lowest ids first (as "Lowest id first (Recommended)" and docs/GDD.md's ascending population ids), and the rest leave service as soldiers sent home do (owner, 9 November 2026: "Sent home (Recommended)"): R-UNIT-005's ledger outcome, by Send home's own code (Demobilization.RecordsHome, written in the capture's one ledger commit, PopulationLedger.Capture, then Demobilization.ReturnArms), with "Nearest city, unemployed (Recommended)", "Settled in (Recommended)", "Keep both (Recommended)", "Back to Arms stock (Recommended)" and "No, records only (Recommended)". Each becomes a free civilian without an assignment, located in the nearest city to the formation's hex by R-UNIT-005's rule (its region plus the city), with reception ticks reception.turns, its qualification and combat XP kept, and its issued military item back in the Arms stock once; after the surrender line come Send home's own Why lines, in Send home's order (owner, 9 November 2026: "Yes, reuse Send home's (Recommended)"): its people line (key sent_home, "Unit N is sent home: its P people are civilians in now", P the leftover records times people_per_unit), then, only when Arms go back, its Arms line ("Yes, one Why line (Recommended)", key sent_home_arms_returned). No aggregate city field is written. Every check comes before the first write: the captives and the leftovers change in one validated ledger commit, so a leftover location the ledger does not bind, like an unbound captor, refuses the surrender with nothing written; the Arms go back after that commit. With fewer records than its strength, all of them are captured. On a ledger game SoldiersHeld grows by the number of records captured, not by the unit's strength, so the check above still holds; without a ledger it grows by the strength, as before. For a formation whose records do not match its strength, the count held, and with it the Money cost per population unit held, therefore follows the records captured. A surrender to a side that is not a bound captor, or made while the held count already disagrees with the Captured records, is refused: Take throws before it writes anything. Called from combat, the throw comes after the engagement's draw and losses and after any earlier engagement of the same command or turn; on a direct call to Combat.Strike or Combat.Resolve those writes stay on the state object given. In the game that object is Simulation.Step's clone (sim/Statecraft.Sim/Simulation.cs:52; combat's callers are Simulation.cs:311, a chosen event option's attack in EventSystem.Choose at EventSystem.cs:347, and TurnResolver.cs:21), so the whole command's or turn's result is discarded and the caller's state is unchanged. Since #1725 PR 6, Step takes a dark start's Historical ledger (Simulation.cs:49, R-POP-002), so a player's attack order runs Combat.Strike on that clone (Simulation.cs:311) and can reach the refusal there; if no order refuses, the turn then refuses the Historical ledger at TurnResolver.cs:16, before its step-2 Combat.Resolve. Either way the throw ends the Step and the caller's state is unchanged. An event option's attack is not reached on a ledger game: EventSystem.Choose refuses any initialized ledger at its entry (EventSystem.cs:207), and a waiting card is refused at EventSystem.Delegate's entry (EventSystem.cs:520). So in the game only a player's attack order reaches the refusal in combat, and otherwise only a direct Combat.Strike or Combat.Resolve does (LedgerPrisonersTests.ARefusedCaptureInCombatThrowsAndAPriorCloneIsUnchanged). Since #2473, Prisoners.Return runs on a Historical ledger (a synthetic or Open one is still refused) and, after the same count check and before any other write, returns the Captured records that the front's sides hold, in ascending population id, in one validated ledger commit, by Send home's own code (Demobilization.RecordsHome) (owner, 9 November 2026, recorded on #2473). Each keeps its population id and becomes a free civilian without an assignment, captor or home. "It's gone (Recommended)": its issued military item is removed from the record and nothing is added to Arms (LedgerPrisonersTests.AReturnedCaptivesIssuedItemIsGoneAndNoArmsAreAdded). "Yes, settled in (Recommended)": its reception ticks are reception.turns, as for soldiers sent home (LedgerPrisonersTests.AReturnedCaptiveIsASettledInFreeCivilianUnderTheSameId, which also pins that another front's captives stay captured and that a second exchange changes nothing). "Keep both (Recommended)": its qualification and combat XP stay (LedgerPrisonersTests.AReturnedCaptiveKeepsQualificationAndCombatExperience). "Nearest to region label (Recommended)": it goes to the nearest of the twelve census cities to its home region's label place (regions.yaml capital_location), by Send home's hex distance and city-id tie (Demobilization.Nearest), located in that city's region plus the city (LedgerPrisonersTests.AReturnedCaptiveGoesToTheCityNearestTheirHomeRegionsLabelPlace, LedgerPrisonersTests.AReturnedCaptiveTiedBetweenTwoCitiesGoesToTheLowerCityId). No score and no Arms line are written for them; the front's SoldiersHeld count goes as before, so the count still equals the Captured records. In the game the exchange card does not reach a ledger game yet: EventSystem.Choose refuses any initialized ledger at its entry and TurnResolver.cs:16 holds the dark turn. Since #1706 the engagement's losses have already removed the surrendering formation's lost records as combat losses (R-COMBAT-006's ledger paragraph).

Under gdd-0.3 (#1710), the prisoners in these pins come from a real surrender (R-POW-001). PrisonersGdd03Tests.PrisonersEatAndAFoodShortfallCostsCohesion pins R-POW-002, PrisonersGdd03Tests.PrisonersAreNeverAResource pins R-POW-003, and PrisonersGdd03Tests.OwnSoldiersAreCapturedCostMoneyAndComeHome pins R-POW-005. The real per-front exchange cards wait on #751, so the return and homecoming pins use the synthetic card until those cards land. On the dark start, LedgerPrisonersTests pins R-POW-001, R-POW-002 and R-POW-005 against the aggregate start with the same counts.

Objectives

R-OBJ-001: An objective (objectives.yaml) names its condition, completes_when: { site_built: <site> } (R-OBJ-004's absorbed has no site, so the site rules here apply only to site_built), a designated site of kind yishuv whose founded date must equal the objective's reference_date. In the last step of each GDD turn (spec §2.3, step 11; D-A6), in objectives.yaml's order (R-OBJ-007), an objective whose site is built and that is not yet recorded is recorded once, with the turn's end date (the site's BuiltOn). Recording it again, or in the legacy loop, changes nothing; a stopped build completes nothing.

R-OBJ-002: Recording a standalone objective adds objective_points + min(objective_bonus_cap, objective_bonus_per_month * monthsEarly) to the score, once. monthsEarly is max(0, monthsFromHistory), and monthsFromHistory counts calendar months from the completion's month to the reference date's month, (y2 - y1) * 12 + (m2 - m1): 0 in the same month ("on time, as in history"), negative when late (owner's card wording, #379, 14 October 2026). Late costs nothing. The owner set 100 points (D-B6) and +5 a month capped at +60 (D-B2) on 14 October 2026 (#456); they are balance values and stay tunable, as rules.yaml design values (R-LOOP-005). R-OBJ-006 says when a multipart objective receives the same base. The rules are those of the date the turn starts from (docs/determinism.md, "Rules per era"), so a turn that ends on the next era's first day scores with the era it started in.

R-OBJ-003: An event with objective: is that objective's card. It is eligible only once its objective is recorded, so it fires once, at the end of the turn the objective is completed in, and never in the legacy loop. It takes no soft-event slot, like a Deport card (owner, 4 November 2026, #2605: "Always that turn (Recommended)"). That holds on the turn that ends the game too (the dated end of R-LOOP-004, or a defeat): its objective cards, and the card of any objective recorded earlier that has not fired yet, fire then, without the cap, and no other event does; and Unity shows them before the end screen (owner, 4 November 2026, #2536: "Show the card (Recommended)"). Answering one there is no order and changes nothing: the score is the one step 11 recorded. A card has exactly one option, an objective and its card name each other, and the card waits for nothing else: its window runs from the game's start to its end (or its era's, since an event may not span an era boundary), with no requires or conditions (the loader checks all three, #2804). An unanswered card resolves to its option without drawing from the RNG.

R-OBJ-004: An objective may instead name completes_when: { absorbed: { population_at_least, overflow_at_most, employed_at_least } } (D-B4; owner, 18 October 2026, #738, option (a)). It holds at step 11 once the population (GameState.TotalPopulation, the national count as step 11 reads it, before the turn's population step, annexed residents included, R-CITY-008) is at or above population_at_least, and then the cities' overflow (R-HOUSE-002), summed, is at most overflow_at_most of their housed people (R-HOUSE-003), summed, and their employed, summed, are at least employed_at_least of their civilians and craftsmen, summed (as R-CITY-005 counts them; owner, 3 November 2026, #1714: "Both rulesets (Recommended)", so gdd-0.1 and gdd-0.3 alike). Refugees waiting do not count: each turn's arrivals (step 9) wait at step 11, after housing (step 8). The gate keeps the census start values (41% employed, no overflow) from completing it on the first turn. Sums are integers, compared with the decimal share, so city order cannot change the result. An objective names exactly one kind; all three keys are required (a missing key is never read as 0), the shares are from 0 to 1 and the population above 0 (the loader). Recorded and scored as R-OBJ-001 and R-OBJ-002 say; its card follows R-OBJ-003. The shipped milestone, 1,000,000, stays Q-ABS-1's sourced number for the Jewish population; since the total counts annexed residents (owner, 4 November 2026, "Count them everywhere", #2620), a game that annexes a stayed town may reach it earlier.

For R-OBJ-004, the preceding active gdd-0.1 rule remains unchanged. The gdd-0.3 addition is an inactive, synthetic-only boundary: Gdd03SyntheticAbsorptionBoundary.HoldsForSyntheticSettledFacts accepts supplied represented own units, a canonical refugee-awaiting-legal-bed flag, ordinary-housing overflow/demand and employed/all-free-civilian counts; it neither produces nor classifies them (sim/Statecraft.Sim/Gdd03SyntheticAbsorptionBoundary.cs:6-25@1e4b8a248). Source: docs/gdd/objective-card.md:19@605277dfa511dc512562ececeb273b8393b21d35, "Empty denominators mean no overflow/full employment." and "Reference remains 20 December 1949"; docs/gdd/balance.md:195-196 at that source revision gives the exceeded population target (superseded below: at or above, #1714) and inclusive percentage limits.

The synthetic boundary reads the supplied catalogue's Agreed population scale, target and percentages. With the current scale and thresholds, 99 units fail and 100 (exactly the target) pass: the target is reached at or above it, in both rulesets (owner, 9 November 2026, #1714: "At or above, both (Recommended)"); overflow exactly 10% and employment exactly 40% pass. The waiting flag does not block the result (see the change below); the tests supply false for a legal reception bed and true for a port queue or refugee evacuation transit, and both pass when the other gates pass, without implementing that classification. Each valid 0/0 pair passes its ratio subcondition, but cannot bypass population. Exact BigInteger cross-products avoid rounded shares. Null/wrong-identity catalogues, missing, Deferred, non-Agreed or malformed quantities and negative/non-subset counts throw before any failing gate can hide them; the population scale retains the positive Int32 bound, the target must be positive and whole, and percentage limits must be whole numbers in 0..100 (sim/Statecraft.Sim/Gdd03SyntheticAbsorptionBoundary.cs:27-65@1e4b8a248; sim/Statecraft.Sim/Content.cs:1270-1278@1e4b8a248; boundary tests in sim/Statecraft.Sim.Tests/Gdd03SyntheticAbsorptionBoundaryTests.cs:12-108@1e4b8a248, scale/threshold/exact-product fixtures at :114-190 at the same head).

Then changed (1 Nov 2026, #1696): the owner chose the option labelled "No reception-bed layer (Recommended)", so in the MVP a waiting flag does not block the result; option (a) (#738) has no waiting clause, and the live R-OBJ-004 check reads population, overflow and employment only (docs/gdd/objective-card.md). Since #2514 Gdd03SyntheticAbsorptionBoundary keeps the flag as a parameter and ignores it.

No live completion or completion date is recorded. The #1714 status comment of 27 October 2026 says, "It does not yet compute actual regional totals or award the objective." Represented population (#1689), regional aggregation and approved locations (#1691; #1690 closed on 2 November 2026, its post-MVP regional pools moved to #2233), shared employment (#1692/#1702, also to feed #1715), canonical waiting (#1696), housing (#1703), and monthly-versus-end-state sampling (#1716) remain pending integration decisions/evidence, not choices made by this helper (sim/Statecraft.Sim/Gdd03SyntheticAbsorptionBoundary.cs:6-17@1e4b8a248; #1714 status comment, "Human check pending: unresolved rules for combining totals and when to count people"). The 20 December 1949 reference is retained; this slice does not select when to award it.

R-OBJ-005: An objective part (part_of) may name completes_when: { corridor_complete: <corridor> } or completes_when: { sites_built: [<site>, ...] }. It completes at step 11 of the first GDD turn when every corridor segment (R-COR-003) or every listed yishuv site is built, records once with that turn's end date and earns min(objective_bonus_cap, objective_bonus_per_month * monthsEarly) once. Every part has a sourced reference_date inside the MVP and its own card under R-OBJ-003. The infrastructure objective's parts are railway, burma_road and yishuvim (Q-OBJ-1 to Q-OBJ-3, owner, 17 October 2026, #782; #791).

R-OBJ-006: Parts with the same part_of are one objective for objective_points: its base points are added exactly once, when the last part records. Earlier parts add only their R-OBJ-005 timing bonus. Rechecking any recorded part changes neither the records nor the score (Q-OBJ-1, owner, 17 October 2026, #782).

R-OBJ-007: Objectives keep objectives.yaml's order: step 11 records them in it, and when two or more objectives' cards fire on the same turn they fire in it (Q-CARD-2, owner, 17 October 2026, #784). Among the events eligible that turn, an objective's card sorts with the others by from as R-EVT-001 orders them, after the other events of the same from (our tiebreak, not part of Q-CARD-2).

R-SCORE-001: The score is the sum of the parts Q-SCORE-1 names (owner, 17 October 2026, #784), and nothing else: objective base points (R-OBJ-002, R-OBJ-006), their timing bonuses (R-OBJ-002, R-OBJ-005) and points for prisoners returned (R-POW-004). The state keeps each part as it scores, so a later reading never rebuilds one from rules that may have changed since (a turn scores with the rules of the date it starts from); ScoreParts.For reads them, and they always sum to GameState.Score.

Money

R-MONEY-001: The GDD's Money is the Treasury (model.yaml money treasury, in IL millions): the player sees it as Money, כסף in Hebrew, on the top bar, the Budget screen and in the causes (spec §1.2, "Treasury becomes Money"; owner, 15 October 2026, Q-MONEY-4 and Q-HUD-3). Debt stays in the sim and the state hash under Money. Foreign currency also remains state, but it is player-facing: R-MONEY-004 shows the reserve on the Defence screen, permits conversion into Money, and keeps the top bar's Foreign currency meter (owner, 2 November 2026: "Spend the dollars (Recommended)" and "Keep it"). The sourced 1951 currency crisis still happens (D-C3, Q-MONEY-1). The MVP keeps both abstractions approved on #884: the per-head tax represents ordinary revenue, while jobs pay the player-sensitive income in R-CITY-002. The rates remain tax_per_capita_per_year 0.0000545 IL million per person-year, income_per_job_per_year 0.06 IL million per employed 1,000-person unit-year, and spending_per_capita_per_year 0.0001036 IL million per person-year, all labelled design values in rules.yaml (owner, 18 October 2026, option 3 and the per-head-tax draft). The same rule applies under gdd-0.3, which shares this economy (D1, owner, 1 November 2026, "Approve all six (Recommended)"; #1702); Amir's jobs-only income and import-only credit are post-MVP. It runs on a game with an initialized Historical ledger (the dark start, R-POP-002; #1725 PR 7), where a head is a living ledger record times people_per_unit (GameState.TotalPopulation, R-CITY-008), and it changes no ledger record; a synthetic or Open ledger is still refused. The tax base (owner, 8 November 2026: "Everyone alive (Recommended)"): every living record pays the per-head tax, whatever its status. The spending base (owner, 8 November 2026: "Same as tax (Recommended)"): spending counts the same people as the tax. The foreign-currency imports base (owner, 8 November 2026: "Everyone alive (Recommended)"): every living record counts toward the per-head imports. The start level (owner, 8 November 2026: "Accept ledger count (Recommended)"): the flows start from the ledger's 552 records, 552,000 people, 15% below the aggregate start's 649,400, and the rates are not retuned (LedgerEconomyTests).

R-MONEY-002: In step 7, after Food consumption and imports, the Treasury pays each existing civilian kind's annual upkeep for the turn's share of a year; an unfinished project pays none, and R-MONEY-003 covers a shortfall. In the MVP, cities, completed permanent-housing projects, completed industry projects, completed transit camps, military camps and Hakhsharot each have an explicit upkeep of IL 0m per year in rules.yaml. These are design values, not historical claims. Phase 5 decides military-unit upkeep separately (#742; owner, 18 October 2026, #884 option 3). It applies under gdd-0.3 too, and the six rates stay IL 0m (the owner's 30 October 2026 "Keep ruling" on UPKEEP; D1; #1702). It runs on a game with an initialized Historical ledger (the dark start, R-POP-002; #1725 PR 7) and pays what the aggregate game with the same places pays: it counts places, not people, and changes no ledger record.

R-MONEY-003: A GDD cost above the Treasury is paid, and the shortfall moves to debt, as map actions do: the Treasury stops at 0, and a cause names what the shortfall was borrowed for (Q-MONEY-3, owner decision "Allow borrowing", 15 October 2026, for training and pioneer work; the Phase 2 plan applies it to every GDD cost). It covers the deficit, map actions, training (R-TRAIN-001), pioneer work (R-SITE-001), an event option's Treasury cost and an arms workshop's cycle (R-ARMS-006). The same rule applies under gdd-0.3 (D1; #1702): debt_ceiling is only the economic-defeat threshold (R-END-001), so every cost still borrows past it, except an Arms purchase whose borrowing would take the debt past it, which is refused (R-ARMS-002, #2198), and an arms workshop's cycle whose borrowing would take the debt past it, which waits unpaid that month (R-ARMS-006). A shortfall changes nothing but the Treasury and the debt, and a refused order changes nothing (Gdd03MoneyLedgerTests).

R-MONEY-004: "Spend the dollars" (#813; owner, 2 November 2026: "Spend the dollars (Recommended)", https://github.com/eyalgolan/statecraft-israel/issues/813#issuecomment-6564883926). ConvertDollars(amount), save line dollars N, scenario kind convert_dollars, converts amount US$ millions of foreign currency into Money: the reserves lose amount and the Treasury gains amount times the mvp.il_per_dollar rate in force on the turn's date, with a cause line, so the turn's deficit borrows that much less (R-MONEY-003); no principal is added to Debt (Q5). A turn's conversions run before its Arms purchases, whatever the queue order, so a purchase borrows only what the conversions leave it short (#2206; owner, 2 November 2026: "Conversion counts first (Recommended)", #813). It is a GDD order, refused in the legacy loop. It is refused, state and Attention unchanged, below $1m ("Convert at least $1m"), outside the war, a rung below 4 at the turn's start ("Dollars can be converted only during the war"), when the turn's conversions would pass mvp.dollars_per_turn (4; "That is more dollars than may be converted this month"), when the game's would pass mvp.dollars_total (100, the Export-Import Bank's 1949 credit, FRUS 1949 vol. VI, d407; "That would pass the total of dollars that may be converted"), or when the reserves would go below 0 ("There are not enough dollars in the reserves"). GameState.DollarsConverted keeps the game's total and is hashed as dollars= only when not 0. The rate is sourced and dated: IL 0.333 per $ ($3.00), the effective dollar rate, to 18 September 1949 (the official sterling parity gave IL 0.248 per $, but most dollar transactions ran at 0.333), and IL 0.357 per $ ($2.80) from the devaluation of 19 September 1949 to 14 February 1952 (IMF, Annual Report 1950, Appendix I, pp. 88-89; Federal Reserve Bulletin, November 1949, p. 1336; Michaely, Israel's Foreign Exchange Rate System (1971), p. 10 and Table 2-1; Patinkin, The Israel Economy: The First Decade, Table 43, p. 121 (gathered on #813)). dollars_per_turn and dollars_total are labelled design values, and the Attention cost attention_cost.convert_dollars is 0 (owner, 2 November 2026: "Keep 0 (Recommended)", #2204). The Defence screen makes the order and shows the foreign-currency reserve beside it. With a conversion queued, a new amount in its box replaces the order at once, or withdraws it when the amount would be refused, so the box, its Money line and what End turn converts agree; the reserve and the total converted keep the turn's start until End turn (owner, 2 November 2026: "Box shows queued order (Recommended)", #813). Outside the war the panel stays visible but disabled, with one line saying conversion is for wartime only (owner, 2 November 2026: "Grey it with the reason (Recommended)", #813); the top bar gains nothing and keeps its Foreign currency meter (owner, 2 November 2026: "Keep it", #2204). The same rules apply under gdd-0.1 and gdd-0.3, which share this economy (D1).

The top bar shows the net position, Treasury minus debt (HudModel.From), so it reads below 0 once anything is borrowed.

Food

The national Food pool (D-C2, "National pools", decided 15 October 2026), built on the maintainer's food rule of 16 October 2026 (#382, option (b): "food_supply adds a stock term: it rises and falls with months of cover (new design values k_food, cover_months), rationing and events still act on top, and the unmet share lowers it further").

R-FOOD-001: Food is a national stock of whole units (GameState.Food); one unit is one population unit's food for one month, and there is no storage cap (Q-POOL-2, decided 15 October 2026). A new game starts the stock at cover_months of the start's need, the need being the national head count in population units (people_per_unit), rounded half away from zero; with an initialized gdd-0.3 ledger (the dark start, R-POP-002) the need is the ledger's, one ration for each living own population id (#2470). k_food and cover_months are required in the GDD loop (rules.yaml mvp:, labelled design values); a missing or out-of-range value is a load error. cover_months runs from 1 to 120 (Pools.MaxCoverMonths), so its multiply cannot wrap the stock (#2804); the stock itself still has no cap.

R-FOOD-002: In step 7 the nation eats the turn's need from the stock as far as it goes: the stock stops at 0, and the rest is the unmet share, kept in the state with the need (Q-POOL-5, decided 15 October 2026). Then a monthly import brings the stock back to cover_months of the need, unless foreign currency is exhausted (the fx_crisis situation at the turn's start), when nothing is imported (Q-POOL-4, decided 15 October 2026). Imports are already paid for in foreign currency per head (imports_per_capita_per_year), so they are not charged again. In the GDD loop the crisis cuts the imports, not food_supply directly: fx_crisis_food_per_year acts in the legacy loop only.

For R-FOOD-001 and R-FOOD-002 under gdd-0.3 with an initialized population ledger (#2470), the need is the month's: one ration for each living own population id in the month's frozen opening snapshot (the #1716 synthetic month context, OpeningPeople), whatever its status or assignment (idle, at work, teaching, on a course, travelling, serving, in reception or held by the enemy), each id once and in any storage order, so a ledger change before step 7 (a combat loss, an arrival) does not move that month's need. The live ledger is reconciled first, so an invalid one moves no Food. The settlement after it is the one above: the stock stops at 0, FoodNeed and FoodUnmet record that need, and the import refills to cover_months of it unless the foreign-currency crisis is on. Counting changes no id, status or assignment. Legacy and gdd-0.1 keep the aggregate head count. Counting a captured own soldier is the owner's decision (5 November 2026, the recommended option, recorded on #2678 as "Captured own-population records do eat the national ration"), the reading of docs/gdd/balance.md population.food_per_turn ("Food per own population unit, every status"); the ledger's head count (TotalPeople) already includes them. A game with an initialized Historical ledger (the dark start) has a live month snapshot (#1725 PR 7, Gdd03SyntheticMonthContext.CreateLive): its opening people are the ledger's living own records, in id order, and its facility entries are the ledger's bound Work targets that are authored fields (field/q,r), in ordinal order, each available at the freeze with field.slots capacity; its month ordinal is the turn (GameState.Turn). It is built from the state as it stands, every check before anything is built, and writes nothing (the state hash and the ledger's canonical text are unchanged); a synthetic or Open ledger and an aggregate game are refused, as is a ledger with a Work, Course, Reception or Teaching target that is not a field. Owner, 8 November 2026, "One snapshot per turn (Recommended)": one month snapshot is taken at the start of the turn, after the player's orders, and harvest, Industry income and Food all read it. Owner, 8 November 2026, "Yes, recheck (Recommended)": a consumer rechecks an entry against the live state before using it, so the snapshot keeps each entry's id and opening record and its availability is only an upper bound. Given the live snapshot, this month path counts the same need as given a synthetic one of the same people (LedgerMonthSnapshotTests). No turn takes the snapshot yet: TurnResolver.Resolve still refuses an initialized ledger, so the change that runs a dark turn takes it at Resolve's entry; without a month an initialized ledger is still refused (R-POP-002).

R-FOOD-003: food_supply stays the 0 to 1 level the UI, the situations and the events read (Q-POOL-3, decided 15 October 2026), with a stock term inside its effect sum: k_food * (stock - full) / full - unmet / need, where full is cover_months of the last need. The term is exactly 0 at a full stock, -k_food when the stock is empty, and the unmet share lowers it further; rationing, the other effects and event shifts act on top as before, and the level's clamp to 0..1 bounds the result. The GDD loop requires the food_supply indicator. The term runs on a game with an initialized Historical ledger (the dark start, R-POP-002; #1704, #1725 PR 7) as in the aggregate game: it reads only the stock, the last need and the unmet share, never a record, and the need is already the ledger's (R-FOOD-001), so no ledger record changes; a synthetic or Open ledger is still refused (R-POP-002).

Arms

The national Arms pool (D-G6, decided 15 October 2026: "From events (sourced arms deals) and Money. Training and attacks consume it"). The four numbers are the owner's (18 October 2026, #781: "Approve as drafted"); each is a labelled design value in rules.yaml, and the balance pass (#813) may tune them. An event can move the Arms stock (R-EFX-001, R-EFX-002), and the Arms an attack spends is Phase 5 (Q-COMBAT-4).

R-ARMS-001: Arms is a national stock of whole units (GameState.Arms, hashed as arms= only when not 0); one Arms is the equipment package for one recruit group of people_per_unit (1,000) people. A new game starts the stock at arms_start (12). It never goes below 0. Under gdd-0.3 an arms workshop's completed cycle adds 1 (R-ARMS-006). arms_start, arms_price and arms_per_recruit_group are required in the GDD loop (rules.yaml mvp:); a missing or negative value is a load error.

R-ARMS-002: BuyArms(count) adds count units to the stock at once, with no delivery delay, and the Treasury pays count * arms_price (IL 1m an Arms), borrowing what it lacks (R-MONEY-003). It is refused, state and Attention unchanged, for fewer than one unit, a stock that would pass a whole number's range, or a purchase whose borrowing would take the debt past debt_ceiling (#2198; owner, 2 November 2026: "Refuse buys over ceiling (Recommended)", https://github.com/eyalgolan/statecraft-israel/issues/2198#issuecomment-6564721486), with the reason "Buying that many Arms would take the debt past its ceiling"; the legacy loop refuses it before charging Attention. The debt after is R-MONEY-003's: the debt plus what the Treasury lacks (Debt + max(0, cost - Treasury)). A debt exactly at the ceiling is allowed, and a purchase the Treasury pays in full borrows nothing and is never refused at it, even with the debt already past it. This overrides Q22 for Arms purchases; an arms workshop's cycle also waits rather than borrow past it (R-ARMS-006), and every other cost still borrows past the ceiling (R-MONEY-003). The Defence screen shows the same reason before End turn when End turn would refuse the amount for the ceiling with the turn's orders as they stand: it runs them from the turn's start on a copy, as End turn runs them, with this purchase in place of a queued one or after them, so the dollar conversions and any other spending queued before the purchase count (#2206). The turn's dollar conversions (R-MONEY-004) run before its Arms purchases, whatever the queue order: a conversion queued after the turn's first BuyArms runs just before it, with the other such conversions in their queue order, and every other order keeps its queue place (#2206). Its Attention cost is attention_cost.buy_arms (1), whatever the count. It runs on a game with an initialized Historical ledger (the dark start, R-POP-002; #1725 PR 7) as in the aggregate game, at the same arms_price: a bought Arms joins the national stock and is not an issued item on any record, so no ledger record changes; a synthetic or Open ledger is still refused (R-POP-002).

R-ARMS-003: Recruit training (R-TRAIN-001) takes arms_per_recruit_group (1) Arms for each group when it starts, after its other checks and before it pays. A stock short of that refuses the order with a reason naming the need and the stock, state and Attention unchanged (Q-UNIT-3). A gun or armoured car crew takes its type's training_arms a group instead, by the same rule (R-UNIT-008). Pioneer training takes no Arms. The cohort keeps the Arms it took, and the state hash writes them with it; cancelling the cohort (R-TRAIN-002) returns them to the pool, up to the most the stock can hold, although the Money stays paid (owner, 18 October 2026, #1047: "Refund on cancel"). Replacement groups (R-TRAIN-004) take their course's Arms by the same rule (owner, 23 October 2026, #1510: "Replacements cost Arms").

R-ARMS-004: Under gdd-0.3 only, the arms workshop (projects.yaml arms_workshop, #1701) is a city project kind: StartProject builds it as R-PROJ-001 to R-PROJ-003 build any kind, for IL 15m and 3 turns (owner, 2 November 2026, #1701 OQ2: "IL 15m, 3 turns (Recommended)"; the gdd-0.3 catalogue's arms_workshop cost 15 and work 3). Built, it gives one workshop worker slot (worker_slots: 1, the catalogue's arms_workshop.parts.worker_slots) and adds no housing, condition or job slots: a workshop gives no Industry job (docs/gdd/cities.md, "Arms workshops"). Its row carries rulesets: [gdd-0.3], and the loader leaves it out of every other ruleset's kinds, so a gdd-0.1 game neither lists nor starts it (StartProject refuses it as an unknown kind) and its state does not change; the row still moves the content hash. The loader refuses an unknown ruleset id in rulesets, worker_slots other than 0 or 1, a kind with a worker slot that has capacity, condition or job slots or is not limited to exactly gdd-0.3, and, under gdd-0.3, an arms_workshop whose worker slots differ from the catalogue's. It pays no upkeep: R-MONEY-002 has no rate for it (the catalogue's 0.50 is a Default). Its Hebrew name סדנת נשק remains a draft until a native speaker, or an AI-assisted contributor who discloses it, has read it (name_he_review: agent-checked, #564; owner, 2 November 2026, #1701 OQ8; owner, 4 November 2026: "AI-assisted can clear it").

R-ARMS-005: Under gdd-0.3, AssignWorkers(city, "workshop", n) is a standing order that staffs the city's first n built arms workshops, in ascending project id, and leaves the rest unstaffed (27M-13, owner-approved 1 November 2026: "Any free civilian. The cost tier follows the worker (novice or ordinary / apprentice / expert)."). A newly staffed workshop takes one of the city's civilians and craftsmen who hold no job; the worker stays one of them, with no job, profession or qualification change, is not employed (R-CITY-001) and earns no job income (R-CITY-002). On the aggregate cities every worker is priced at the ordinary or novice tier until the population ledger's tiers are live (owner, 2 November 2026, #1701 OQ6: "Accept all four (Recommended)"; R-POP-004). It is refused, state unchanged, for an unknown city, a negative count, more than the city's built workshops ("Town A has 1 arms workshop"), or more newly staffed workshops than the city's people without a job ("... civilians and craftsmen without a job"); under gdd-0.1 the workshop role is refused as "There is no such role", and the legacy loop refuses the order before charging Attention ("That order is not available in this game", R-CITY-001). While the city's job default is on (R-CITY-009), a newly staffed workshop may take its worker from the default's jobs instead: the order is refused only past the city's civilians and craftsmen less its staffed workshops and builders (R-CITY-011) ("... civilians and craftsmen"), the city's employed falls to at most its civilians and craftsmen less the staffed workshops and builders (a city cause, "Town: N groups in work"), and the default stays on, filling one job fewer (R-CITY-001; #2270 review fix, 2 November 2026). A number of jobs the player set (default off) is never lowered: there the order needs people without a job. A builder (R-CITY-011) is not one of the city's people without a job for this order. A staffed workshop's worker is not idle: Train (R-TRAIN-001) and ReassignPioneers (R-QUAL-005) cannot take them, and the Train panel's limit (Training.Available) counts a pending workshop order's newly staffed workers, never as jobs. An unstaffed workshop keeps its record, paid cost and progress (a pause, R-ARMS-007). There is no limit on workshops per city beyond one project a hex and the starts a turn (R-PROJ-001; #1701 OQ7). GameState.Workshops holds every workshop ever staffed, by project id, and is hashed as wshop[...] only when it has one (R-LOOP-002). The hex panel of a built workshop (#1722, #1701 slice E) queues it: Workshops.Check(city, n, pending) gives the order's own refusal and, with this turn's queue, also counts the city's pending job order and its Train and reassignment orders, so a count it accepts is accepted whichever runs first; the jobs stepper's limit (Cities.MaxWorkers with the queue) counts a pending workshop order's newly staffed workers, as Training.Available does. The panel's next-cycle price is Workshops.NextCyclePrice, what the production step would charge (R-ARMS-006), shown for a workshop never staffed too, at the ordinary or novice tier, and its cycle length is Workshops.CycleMonths. A jobs order past the city's civilians and craftsmen less its staffed workshops is refused with the staffing order's people format, "Town A has N civilians and craftsmen" (#1722 review). On a game with an initialized Historical ledger (the dark start, R-POP-002; #1725 PR 7) the same order runs on the people records, under the owner's rulings (owner, 8 November 2026): "Keep existing order (Rec.)": AssignWorkers(city, "workshop", n) is the order there too, and the order picks the records; every other role is still refused. "Unassigned, lowest id (Rec.)": a workshop no longer staffed first releases its worker record, an unassigned free civilian again in place; then each newly staffed workshop, in ascending project id, takes the city's unassigned free-civilian record (located in the city) with the lowest population id, so the Haifa teacher, the city's last record, is taken last. "Work assignment (Rec.)": that record holds a Work assignment on the target workshop/<project id> in the city's region, bound the first time the workshop is staffed (owner, 8 November 2026, #1725 issuecomment-6725822701: "At first staffing (Recommended)"), in the same all-or-nothing ledger commit as the records; it keeps its status, place, profession, tier and XP (Q55). "Keep tier field (Rec.)": the workshop's WorkerTier is the record's priced tier (R-ARMS-006), with no new field. The order is refused, state unchanged, for an unknown city, a negative count and more than the city's built workshops as above, and for more newly staffed workshops than those records ("Haifa has N free civilians without an assignment"). "Wait for Industry records (Rec.)": no count of the city's people without a job is made, and no city field is read or written, so the job default's path above (a worker taken from the city's jobs) waits for #1725 PR 3's Industry records. A city with builders (R-CITY-011) is refused, state unchanged and nothing thrown, since builders are not records yet (owner, 8 November 2026, #1725 issuecomment-6725822701: "Refuse with a reason (Recommended)"), with "Haifa has builders on city projects, who are not on the population ledger yet"; this internal English text gets its Hebrew with the UI wiring PR. Workshops.Check and the panel's limits are the aggregate game's and are not used there: they still read the city counts, and the workshop UI PR makes them read the records (owner, 8 November 2026, #1725 issuecomment-6725822701: "Fix in the UI PR (Recommended)").

R-ARMS-006: On step 6 of the turn, after the cities' job income, in ascending project id, each staffed arms workshop under gdd-0.3 works this month when its city is not held (R-CITY-010) and one of the city's people without a job is left for it once step 5's builders have taken theirs (R-CITY-011; each workshop of the city takes one, in ascending project id; when none is left for it, it waits, and the Why screen says so). A workshop with no cycle under way first pays for one there (27M-11: "At the monthly production step."), at its worker's tier from the gdd-0.3 catalogue: workshop.cost IL 5m / 4m / 3m (ordinary or novice / apprentice / expert; docs/GDD.md's balance ruling: "workshop.cost is Money per Arms and stays."), beside the IL 1m an Arms of R-ARMS-002 (owner, 2 November 2026, #1701 OQ1: "Keep as ruled"). A short Treasury borrows the rest (R-MONEY-003; 27M-12: "Automatic each month while the workshop is staffed, paid like any GDD cost."), but a cycle whose borrowing would take the debt past debt_ceiling is not paid: the workshop waits that month, the turn report says why (below), and nothing is borrowed (owner, 2 November 2026, #1701 OQ4: "Refuse past the ceiling (Recommended)"; exactly at the ceiling is allowed, and a Treasury that pays in full is never refused, as in R-ARMS-002). The month a cycle is paid is its first staffed month (#1701 OQ3: "Accept all four (Recommended)"). After workshop.staffed_turns (3) staffed months the cycle adds 1 Arms to the stock (R-ARMS-001) and ends, and the next month starts the next one by itself, with no order. When the Arms stock is int.MaxValue, a finished cycle waits silently at its completed, paid progress, keeps its worker, and adds no Arms or Why line. A worker change keeps the cycle's paid cost and progress; only the next cycle is priced at the new worker's tier (27M-13). An unstaffed workshop, or one in a held city, keeps its cycle as it is and is not charged. The turn-report panel lists, in a row of its own after the stocks, each Arms a workshop made ("Workshop in Town: +1 Arms") and each ceiling pause ("Workshop in Town paused: debt ceiling") (owner, 2 November 2026, #1701 A1: "Yes, in the panel (Recommended)", the option "The turn-report panel shows 'Workshop in : +1 Arms' and 'Workshop in paused: debt ceiling'."); a no-worker pause is on the Why screen only. A city's Arms are one line and its ceiling pauses another, each with its count when there are two or more ("Workshops in Town (2): +2 Arms", "Workshops in Town (2) paused: debt ceiling"), in the order each first appears among the turn's causes (owner, 2 November 2026: "Combine per city (Recommended)", the option "'Workshops in Tel Aviv (6): +6 Arms'; pauses likewise combined with a count."). The option gives no wording for one workshop, so a city with one keeps A1's line. The count is the number of the sim's causes; the UI computes nothing else. On a game with an initialized Historical ledger (R-POP-002; #1725 PR 7) a staffed workshop works when its own worker record (R-ARMS-005) is still a free civilian of the workshop's city on its target; a builder never takes that record (R-CITY-011), so no count of the city's people without a job is read (owner, 8 November 2026: "Wait for Industry records (Rec.)"). "Waits, stays staffed (Rec.)": a workshop whose worker is gone waits with the no-worker Why line, stays staffed and is given no other record. "Only craftsmen discount (Rec.)": a new cycle is priced at the record's tier, IL 4m for a craftsman apprentice and IL 3m for an expert, and IL 5m for every other record (profession none, a farmer or pioneer at any tier, a craftsman novice); "Keep tier field (Rec.)": the tier is written to WorkerTier when that cycle is priced, and a cycle under way keeps its paid cost. The held-city pause, the payment, the borrowing, the ceiling, the Arms and the Why lines are as above. A synthetic or Open ledger is still refused.

R-ARMS-007: A workshop's cycle progress is lost and its Money is not refunded on a destructive cancellation, destruction or transfer of the workshop (27M-14, owner-approved 1 November 2026: "Cycle progress is lost and the Money paid is not refunded."). The MVP has none (owner, 2 November 2026, #1701 OQ5: "Accept all four (Recommended)"): a built project cannot be cancelled (R-PROJ-002), a held city pauses (R-CITY-010) and unstaffing pauses (R-ARMS-005), and each keeps the cycle; a later mechanism that destroys or transfers a workshop applies this rule. Workshop work changes no qualification, profession or craftsman count (Q55, owner-approved 1 November 2026: "No. Only the stated apprentice-to-expert progression counts"). On a Historical ledger game workshop months give no record work XP yet (owner, 8 November 2026: "Not in this change (Rec.)"); the craft work-XP producer adds them.

Defeat

The game can be lost (D-B5, decided: "Money and Food exhausted for N turns, or a war loss. Values for the owner. Coalition collapse is off in the MVP"; Phase 7 PR 5, #8). This part covers the economic defeat (R-END-001, R-END-002), which remains in gdd-0.3, and the gdd-0.1 war loss (R-END-003). The defeat texts are decided (Q-DEF-4 (b), owner, 18 October 2026, #784): a short factual line per cause - "Money and Food ran out" and "The enemy holds a city" - plus one sourced history note beside each, which goes on the end screen with #812.

R-END-001: Money is exhausted when the debt is at or above debt_ceiling (Q-DEF-1, owner, 17 October 2026, #784: a debt ceiling; below it borrowing goes on as R-MONEY-003 allows). Food is exhausted when the national stock is 0 (R-FOOD-002; Q-POOL-5, "The stock stops at 0"). One of them alone ends nothing. debt_ceiling is IL 250m, a labelled design value (the recommended first value; the Phase 7 balance pass, PR 12, sets the final number), required in the GDD loop, a finite number >= 0.

R-END-002: In the resolver's step 11, after the objectives and the score (spec §2.3) and before the war loss (R-END-003) and, under gdd-0.3, each front's quiet-turn count (R-WAR-005; the resolver runs it last, after the spec's 11 steps), a turn end with Money and Food both exhausted adds 1 to ExhaustedTurns, and any other turn end sets it back to 0. When it reaches defeat_turns, the game ends with defeat_economy and a cause in the turn's causes, under the game's end (era) like the MVP end. defeat_turns is 3, a labelled design value (Q-DEF-2, owner, 18 October 2026, #784: "3 turns"), required in the GDD loop, at least 1. The step reads the debt as it finds it: the turn's deficit is borrowed after the resolution, so it counts from the next turn's end. A defeat on the game's last turn ends it as a defeat. ExhaustedTurns is in the state hash only when above 0, so a run that never has both exhausted hashes as before; the legacy loop never counts.

R-END-003: Under gdd-0.1, when an enemy military unit (a unit with a side other than the player's; a halutz has no side, R-UNIT-001) holds the hex of one of the player's GDD cities (cities.yaml; the hex in map/settlement_hexes.yaml) at the end of a turn, the game ends at once with defeat_war and a cause in the turn's causes, under the game's end (era). One turn is enough, any of the GDD cities, at any time, truces included (Q-WAR-5; Q-DEF-3, owner, 17 October 2026, #784). The check runs in the resolver's last step after the economic count (R-END-002), so a turn the economic defeat also ends keeps defeat_economy.

Under the explicit gdd-0.3 ruleset, this city-loss check is skipped before scanning cities or units: losing a city or region alone does not set GameOver or emit a defeat_war cause. The existing economic and dated ends retain their order; an already-ended game is never revived. Authority: the owner's two rulings recorded on 25 October 2026 on #1668, “there is an economic defeat” and city or region loss does not end the game. This exception does not implement region transfer, evacuation, recapture or new enemy capture permissions; R-CTRL-006's loss of a player's locality ends no game here either. It changes no legacy or gdd-0.1 rule.

The end report

D-H4 as decided (owner, 17 October 2026, #784: Q-END-1 and Q-END-2 as recommended): #19's legacy report folds into the end screen, which reads one sim helper, EndReport.For (Phase 7 PR 7, #810). This first slice reads only what the state already holds; the score's parts wait for #809.

R-END-005: Once the game has ended, the end report lists the outcome (the GameOver code), the end date, the state's score as it stands, every objective in objectives.yaml order with its historical date and, when completed, the player's date and the months from history (Objectives.MonthsFromHistory), any Deport card that never fired although its place was depopulated (a Deport on the last turn, R-EVT-007), and the two national parameters, Social cohesion and International Support, with no historical value. It is pure: it computes no rule, writes nothing, and two calls give the same report with the state hash unchanged. A game still on has no report.

R-END-006: The report compares one measure with history (Q-END-2): the population at the end (GameState.TotalPopulation, annexed residents included, R-CITY-008) against the latest CBS anchor of the Jewish population in population.yaml dated on or before the end, the anchor labelled as the Jewish population, given with that anchor's date and sources. A game that ends before the first anchor has no comparison. No other value is compared, and none is invented.

Government

D-C5 as decided (12 October 2026): the cabinet and Attention stay, the coalition is optional with no collapse. The coalition votes, the budget shares and the policies keep running and visible as before (Q-GOV-1, decided 15 October 2026, "Take all recommendations", #375). The defeat (D-B5) is built in Phase 7 (Q-GOV-2, decided the same day).

R-GOV-001: In the GDD loop the coalition's stability never ends the game: it still moves toward public approval each turn and still weights the coalition votes, but a value below 0.2 (PoliticsSystem.CollapseBelow) sets no coalition_collapse (D-B5: "Coalition collapse is off in the MVP"). The legacy loop keeps its collapse while the legacy code exists (D-E1).

R-GOV-002: Every GDD order kind has an attention_cost in the real rules (a design value, R-LOOP-005), the unit orders pioneer_work and stop_work cost 0 (D-A6: unit orders are free), and a refused order costs nothing: its Attention is refunded before the next order in the turn is priced. This accounting contract applies to legacy/gdd-0.1. Under gdd-0.3 there is no Attention resource: no refill, charge, refund, affordability query or canonical att field; real order validation/material costs remain, and capture previews express absence (#1688; owner ruling, 25 October 2026, on #1668). For gdd-0.3, current owner: ministers-events-end.

D-H3's first slice in the GDD loop (GDD Phase 7 PR 4, #808): each office's competence scales the GDD systems in its remit. Q-MIN-1 and Q-MIN-3 were decided as recommended, Q-MIN-2 as "Output" (owner, 17 and 18 October 2026, #784): competence scales output and progress a turn, and the base price and stated duration stay unchanged. The system-to-office map is mvp.minister_systems in content/rules.yaml, each entry a labelled design value; the corridors (#870) and military training (#742) join with their own phases. Candidates and appointment stay Cabinet.Appoint's.

R-MIN-001: Each GDD system named in mvp.minister_systems reads its office's cabinet multiplier on the turn's date - the holder's competence factor times the budget factor, EffectsSystem.Multiplier - through EffectsSystem.GddSystemMultiplier (Q-MIN-1). A system the map does not name is neutral (1.0), and a multiplier within 1e-9 of 1.0 reads as exactly 1.0, so float noise never stalls a neutral system. The loader refuses a map entry whose office offices.yaml does not define.

R-MIN-002: The multiplier scales output and progress a turn (Q-MIN-2): a site being built adds progress x m of a turn's work (yishuv and field work, R-SITE-003; a veteran's veteran_progress is the progress), a city project adds 1 x m (housing and industry construction, R-PROJ-003), a corridor segment adds progress x m under Transportation (R-COR-002; 1 or a veteran's veteran_progress), and absorption houses floor(refugees x m) groups a turn (R-CITY-004). The base price and the stated duration stay unchanged. The rounding rule is one: the integer part of the work applies at once, the fraction is carried on the record (R-SITE-003), so no fraction of a turn's work is lost and no whole group is housed early; the hash of a build with no carry is unchanged.

R-MIN-003: When the office does not exist on the date - ended, or not yet created - the system reads the multiplier of the office's named successor, or is neutral (1.0) with none (Q-MIN-3); the absent-holder 0.5 of EffectsSystem.Multiplier stays the rule for an office that exists but has no holder on the date.

National parameters

The GDD's Social cohesion and International Support (D-C1, decided 15 October 2026: "0–1 indicators through the effects web. Cohesion replaces morale on screen. Thresholds as situations"). Their numbers are labelled design values in model.yaml (Reading 1).

R-NAT-001: Social cohesion is the 0 to 1 morale indicator, shown to the player as Social cohesion; its id stays, so no event changes (Reading 2). Its sources are the existing homes, rationing and security links, the cities' employed share (R-CITY-005) and food_supply, so a Food shortage lowers it (food_lifts_cohesion; Q-NAT-2). Approval keeps its weights, which now read Social cohesion through morale (Q-NAT-5).

R-NAT-002: "Low social cohesion" (low_cohesion) is a situation on morale: on below 0.30, off above 0.45, as food_crisis. It has no mechanical effect of its own; events can read it through their indicator conditions (Q-NAT-3).

R-NAT-003: International Support is the 0 to 1 intl_standing indicator, shown as International support; event effects on intl_standing move it (Q-NAT-4).

R-NAT-004: "Low international support" (low_intl_support) is a situation on intl_standing with the same bands and, likewise, no effect of its own (Q-NAT-3).

R-EFX-001: In legacy and gdd-0.1, an option's effect may target an indicator id from model.yaml, treasury or fx_reserves (game money), or food, the national Food stock: the effect moves the stock by whole units (the delta rounds half away from zero) and the stock never drops below 0, so a cut larger than the stock applies only to what is there. Any other target is a load error naming file and field. The Arms stock becomes a target with its pool (#781): arms moves it the same way.

Under gdd-0.3 (#1718, first slice landed in PR #1772 on 26 October 2026): only the active event catalogue is simulated. content/gdd03-events.yaml is required, read and hashed only for this identity; it explicitly excludes whole cards by id with a reason. Unknown or duplicate exclusions and stranded requirements, variants, rejoin edges or objective cards are errors. All source YAML is still parsed strictly. No legacy card is excluded, and no source bytes or reviews changed for it (eight until the owner's ruling of 2 November 2026, "Bring back 3, design bonds (Recommended)", restored clothing_rationing_1950, reparations_claim and reparations_vote, #2224; five until israel_bonds came back with its Debt, R-EFX-009, #2457; four until the owner's ruling of 4 November 2026, "allow cards to effect food supply and housing readout, get these cards into the game turned on", turned on austerity_1949, forex_crisis_1951, iraqi_exodus and maabarot_policy, #2483). There are no replacement amounts. Approved indicator writes and explicit treasury, food, arms effects retain finite checks, stock rounding and bounds. Attention and city-loss/defeat command targets are rejected before compatibility-indicator acceptance. Under the owner's ruling of 4 November 2026 (#2483) events write food_supply and the national housing indicator as under gdd-0.1: the effect moves the readout through its clamped event shift. Under decision D1 and answers Q1 and Q2 (owner, 1 November 2026, "Approve all six (Recommended)" on #1668) the passive budget, the monthly Food refill and the foreign-currency ledger stay, so fx_reserves and the budget and Food-refill spellings load under gdd-0.3 exactly as under gdd-0.1: fx_reserves moves the ledger, and a budget or refill spelling is accepted only where model.yaml defines it, otherwise it is an unknown target. The rationing and maabarot policy writes of austerity_1949 and maabarot_policy load and apply as under gdd-0.1 (#2483; a maintainer reading of "turned on", not an owner ruling). Structured grant_units, city_effects and unlock_projects await approved population/facility accounting and are errors in this slice. Explicit Treasury costs continue borrowing with the existing cause under R-MONEY-003; the owner's 26 October 2026 ruling on #1718 overrides the proposal's no-borrowing text. No transaction refusal or new debt limit is introduced. Every effect field of every event option, with its gdd-0.3 target (the same, or legacy rulesets only for an excluded card), its delta and its review state, is listed in docs/gdd03-event-inventory.md; Gdd03EventInventoryTests regenerates that table from the loader and fails when the file no longer matches the content (#1718).

R-EFX-002: In every ruleset, an option of an event tagged holocaust, terror, massacre or war-casualties cannot raise treasury, fx_reserves, food, arms, food_supply, morale or intl_standing (Q-EFX-2, decided 15 October 2026: the no-gain rule covers all five resources, with the legacy indicators standing for Food, Social cohesion and International Support; food is Food's own target, and Arms joins the rule with its pool, #781). A cost stays allowed. Under gdd-0.3, R-EFX-001 first rejects retired compatibility targets; the remaining Money, Food, Arms and indicator effects keep these no-gain checks. Availability changes do not reset sensitivity tags, review status or human sign-off, and no alias grants an exception (R-EFX-005/007).

R-EFX-007: An option of an event tagged displacement cannot raise Money (treasury or fx_reserves), Food (food or food_supply), Arms (arms), Social cohesion (morale), International Support (intl_standing) or score, with no exception (Q-DEP-6(a), owner ruling of 18 October 2026). Score is not an event-effect target yet and remains an unknown target under R-EFX-001; when that target exists, this rule covers it. The tag also receives R-EFX-005's ban on unit grants and positive city effects.

D-H2 in the GDD loop (GDD Phase 7 PR 2, #806). The owner took the plan's recommendations for Q-HFX-1 to Q-HFX-3 on 17 October 2026 (#784): grants of units, a city's housing capacity, condition and job slots, and project unlock, each validated at load; a grant is free and stands at the military camp or at the site the event names (in this slice, a city the event names: a site waits until sites carry a date, R-EFX-006); the no-reward rule covers them. The first slice built unit grants and the housing capacity and job slots; the second adds the housing condition (#855); the third adds project unlock (R-EFX-008). The owner ruled on 19 October 2026 (#1095) that no current projects.yaml kind gets an additional lock, so none is locked.

R-EFX-003: An option's grant_units entry { type, n, at } adds n (at least 1) units of a unit_types.yaml type of kind unit, free, taking the next ids in turn. Military types are refused: this schema does not specify military strength or a sourced formation, and must not create a default-side, zero-strength soldier. Recruits enter through training (R-UNIT-003); a military event grant needs a separate schema decision (owner, 19 October 2026, https://github.com/eyalgolan/statecraft-israel/pull/1130#issuecomment-6269103829). Granted units stand at the settlement hex of the city at names or, with no at, at the settlement hex of the one city whose facilities include military_camp (Q-HFX-2). With no such single city, at is required. A site is not yet a target (R-EFX-006).

R-EFX-004: An option's city_effects entry { city, field, delta } changes only that city's housing_capacity or job_slots, by a whole, non-zero delta, or its housing condition by a non-zero delta from -1 to 1. Housing capacity never drops below 0, and job slots never drop below the city's employed, so a larger cut applies only to what is free. Capacity may fall below the city's civilians, as overcrowding already can. The condition stays from 0 to 1 (R-HOUSE-001). The other fields of the city, its condition included, do not change. The cause carries the change that applied.

R-EFX-005: An option of an event carrying any sensitive tag (holocaust, terror, massacre, war-casualties or displacement) carries no unit grant and no city_effects gain, whether capacity, slots or condition, and no project unlock (Q-HFX-3; extends R-EFX-002 and R-EFX-007). A cut stays allowed.

R-GATE-001: An event tagged displacement must have status: sensitivity-reviewed before a shareable build. The release gate lists every such event whose status is not reviewed and refuses the build (Q-GATE-1(c), owner ruling of 18 October 2026).

R-GATE-002: A shareable build is refused unless every settlement-control string (the capture and end-of-occupation dialogs, the map-state labels, the info-panel state lines and the International Support count line), in English and Hebrew, matches settlement-control-strings-review.yaml and that file's status is sensitivity-reviewed (Q-GATE-1 (c), owner, 18 October 2026).

R-EFX-006: The structured effects apply after the numeric effects, grants first, then city effects, then project unlocks, each in the option's list order. A target that does not exist on the date (an unbuilt site, a city not yet in the game) is refused at load with file and field, never skipped at run time; so is an unknown type, city, field or project kind. Every cities.yaml city is in the game from the first turn. A site has no window and starts unbuilt, so a site id as at is refused until sites carry a date. The legacy loop, which has no cities or units, refuses them. The preview (R-EVT-005) lists them in the same order.

R-EFX-008: A projects.yaml kind may say locked: true (default false). StartProject refuses a locked kind, like the other R-PROJ-001 refusals, until an option whose unlock_projects lists it has applied; from then on it is an ordinary kind for the rest of the game. The unlocked kinds are game state, hashed as unlocked[...] only when one is unlocked (R-LOOP-002). The loader refuses an unlock_projects entry that names an unknown kind or one that is not locked, since it would do nothing.

R-EFX-009: Under gdd-0.3 only, an option's debt_dollars (US$ millions) also adds to Debt, in IL millions: that amount times the mvp.il_per_dollar rate in force on the event's date (the day EventSystem.FillDate gives, inside the window), scaled by a delegation's share as the numeric effects are (R-EVT-003). It adds no interest and no repayment schedule (GDD Q5). Its line is an event_amount line on debt, and it is listed in EventSystem.OptionEffects after the numeric effects (R-EVT-005). The Debt it adds has no Treasury line, so EconomySystem.MovesMoney counts that line as a Money (net) cause, with its sign reversed (EconomySystem.MoneyDelta). The loader refuses a value that is not a finite number above 0 in every ruleset, since every ruleset reads the same cards; legacy and gdd-0.1 apply nothing. The owner's rulings of 2 November 2026 on #2224 ("Add equal Debt", "Rate in force on the date (Recommended)", "No Debt for donations; use 52 (Recommended)") give israel_bonds/launch debt_dollars: 52 (IL 18.564m at 0.357 on 10 May 1951), and donations_only none (#2457). Tests: Gdd03EventRulesTests.DebtDollarsAddDebtAtTheEventsDateRateUnderGdd03Only, DelegatedHalfShareHalvesTheDebt, DebtDollarsMustBeAFinitePositiveAmount, DebtDollarsThatIsNotANumberIsRefused, IsraelBondsLaunchAddsEqualDebtAndDonationsAddNone, RestoredCardsLoadWithLegacyEffectsAndFireOnTheirTurns, Gdd03MoneyLedgerTests.EveryTurnOfAWholeGdd03GameReconciles.

Events

D-H1 in the GDD loop (GDD Phase 7 PR 1, #805). The owner took the plan's recommendations for Q-EVT-1 to Q-EVT-4 on 17 October 2026 (#784). As the plan's Decided rows record them (docs/superpowers/plans/2026-10-15-gdd-phase-7.md), that keeps windows, hard anchors and the cap of 3 soft events a turn, keeps delegation, adds no delayed effects in the MVP, and shows every effect of every option from the sim's preview. These rules state what the code already does; no behaviour changed. On 18 October 2026 the owner added loyalty to delegation (#107: "Yes: minister loyalty changes a delegated event's outcome, with labelled values in the PR"); R-EVT-003 states it, with two design values (delegation_loyalty_floor, disloyal_effect_share).

R-EVT-001: An event fires on the first turn whose span overlaps its window [from, to) (or at the new game, when the start date is in it), once its requires is met and its conditions hold; an objective's card also waits for its objective (R-OBJ-003). Every eligible hard anchor fires. At most 3 soft events fire a turn (EventSystem.MaxSoftEventsPerTurn), apart from the cards R-OBJ-003 and R-EVT-007 say take no slot; ordinary soft events take the slots the hard anchors leave, max(0, 3 - hard anchors), and the rest wait for the first later turn with a free slot, even after their window has ended (Q-CUT-6; owner, 4 November 2026, #2605: "Show it late (Recommended)"). A waiting event's normal eligibility gates, including its conditions, are evaluated again when it can fire. R-ENEMY-007's cards are the exceptions: an operation card never first fires after its window has closed, and the wait for a slot does not keep it past its window; a card with follows may fire after its window, on the turn after its requires are met; and a card with armistice_front never first fires after that front's armistice is signed.

R-EVT-002: The soft events take the free slots in order: an early armistice first (front_quiet, R-WAR-008; Q-EVT-1, Q-OBJ-1), then consequence cards (R-ALT-008), then the rest; within each group by from (an objective's card after the others of its from, R-OBJ-007), then order (R-EVT-006), then id (ordinal). So two fronts that turn quiet on the same turn come in a stable order, and the cap of 3 still holds. With no early armistice eligible the order is unchanged.

R-EVT-003: An event the player leaves unanswered at End Turn is delegated to its office's holder (Q-EVT-2): with probability competence/10 the holder picks the recommended option (else the first), otherwise a random option. In Historical mode a hard anchor takes its historical option, and an objective's card its one option, with no RNG draw. A holder whose loyalty is below delegation_loyalty_floor (5) applies only disloyal_effect_share (0.5) of each numeric effect of the option drawn; the grants, city effects, policies and security level apply whole, and so do a hard anchor's historical option and an objective's card. Why: competence decides which option a minister picks, loyalty how hard he works to carry out a decision the player left to him (#107, owner, 18 October 2026: "minister loyalty changes a delegated event's outcome"). Both are design values (R-LOOP-005), required by the loader: the floor 0 to 10, the share 0 to 1. Every minister in people.yaml is rated 5 or more today (5 is the deputies' neutral rating), so the rule changes no shipped game until a rating goes below the floor.

R-EVT-004: An option's effects apply when it is chosen or delegated, on that turn, and never later; an option has no delayed form, and the loader refuses an unknown key such as delay (Q-EVT-3).

R-EVT-005: One sim helper, EventSystem.OptionEffects, lists every effect of an option in the order choosing it applies them: the numeric effects, the unit grants, city effects and project unlocks (R-EFX-006), the policies it sets (by id), then the security level, then the way back to the record (rejoin, R-ALT-009), each with the option's own amount (Q-EVT-4). The event card will show this list (GDD Phase 7 PR 8), so no number will be computed in Unity.

R-EVT-006: Among the events eligible on a turn with the same from (and, for objective cards, the same place in objectives.yaml, R-OBJ-007), an event's order (default 0) decides before the id (ordinal): a lower order fires first, takes a soft slot first and is shown first. With every order 0 the order is R-EVT-002's. degania_defence has order: 1, so One Army (idf_founding) opens the game and the defence of Degania comes second on the same turn (owner, 24 October 2026, #1568: "One Army first"). armistice_jordan has order: 1, so on 15 March 1949 Lebanon's armistice card comes before Jordan's, as Lebanon signed first (23 March 1949; Jordan on 3 April) (#2551).

R-EVT-007: An event with deported: <target> is Deport's follow-up card for that settlement target (GDD Phase 6 PR 9, #801; Q-CARD-1 (c), owner, 18 October 2026, #783). Besides R-EVT-001, it fires only once the target's record is depopulated, which only the player's Deport does (R-CTRL-018, R-CTRL-020), so it fires at the event step of the turn that Deport resolves and never after Raid or Occupy, and a Deport of another place does not fire it; only the GDD loop records settlements, so the legacy loop never fires one. It takes no soft slot (R-EVT-001's cap of 3 does not count it), so hard anchors or other cards on that turn never push it to a later one. On the turn that ends the game no event fires but the unfired cards of recorded objectives (R-LOOP-004, R-OBJ-003), so a Deport on that last turn raises no card; the end report lists that card instead, and the end screen shows its title and text after the objectives (owner, 2 November 2026, #2197: "Show on the end screen (Recommended)"). A card that fired is not listed again. It has exactly one option, and, like an objective's card, it resolves to that option without an RNG draw when left unanswered (R-EVT-003). The loader refuses one whose target is not in settlement_targets.yaml, whose target's residents fled (Deport is refused there, so the card could never fire), that lacks the displacement tag (so the release gate holds it, R-GATE-001), that has more than one option, a to (a closed window would miss a later Deport) or prisoners, that is not soft, that opens after the game's start, that waits for anything else (conditions, requires, site, objective, a kind other than historical, variants, front_quiet), or that is a second card for the same place. The shipped cards, deport_acre and deport_nazareth, are sensitivity-reviewed (owner, 2 November 2026, #2197: "Approve both cards"); Deport stays out of play (mvp.deport_in_play: false).

Alternate timelines

The alternate-timeline mechanism (the plan; the owner's answers, 17 October 2026, #945).

R-ALT-001: An event's requires is a list, and the event is eligible only when every entry holds. An entry { event, option } holds when that event was resolved (chosen or delegated) with that option. A list names each event at most once.

R-ALT-002: An entry { event, diverged: true } holds when that event was resolved with any option except its historical one. An event with no historical option cannot diverge, and the loader refuses diverged on it.

R-ALT-003: A non-historical option may carry alternative: { basis, by, date, sources }. basis is proposed or debated (a body or office, in words, is recorded putting it forward on that date; at least one https source with a checked date) or design (nobody is recorded proposing it: a design assumption, which names nobody and no date; sources optional). The historical option carries none.

R-ALT-004: Every non-historical option of an event that has a historical option carries an alternative record (R-ALT-003).

R-ALT-005: An event of kind: consequence is soft, has no historical option, has at least one requirement naming a road not taken (diverged: true, a non-historical option, or an option of another consequence card), carries an alternative record on every option, and lies at most 2 cards deep from the record (EventRules.MaxConsequenceDepth, a labelled design value, Q-ALT-5). It is never an objective's card, and only a consequence card may require one.

R-ALT-006: A consequence card may follow only an option whose alternative.basis is proposed or debated; with diverged: true, every non-historical option of that event must be (the Ferguson test, Q-ALT-3).

R-ALT-007: A consequence card carries every sensitive tag (case aside) of each event it requires, so the no-gain rule (R-EFX-002, R-EFX-005) covers it.

R-ALT-008: Consequence cards take a turn's soft slots before the other soft events, then by from (an objective's card after the others of its from, R-OBJ-007), order (default 0, R-EVT-006) and id (Q-ALT-14). An early front_quiet armistice (R-WAR-008, R-EVT-002) comes before them. With no consequence card eligible the order is unchanged.

R-ALT-009: An option of a consequence card may rejoin an event whose road the card follows (and which has a historical option). Once it is chosen, a requirement on that event's historical option holds, so the record's later cards fire again if their windows are still open. The event's own chosen option is not rewritten, and diverged: true on it still holds.

R-ALT-010: An event of kind: reconvergence has a historical option and may carry variants. The first variant whose every when entry holds supplies the card's title (if it gives one) and text; options and effects never vary. Each variant follows a road not taken, carries sources, has its placeholders checked over the event's window, and names only roads whose windows close on or before the event's from (for a road's historical option, also every consequence card that can rejoin it, R-ALT-009). Only a reconvergence event carries variants.

R-ALT-011: A shareable build refuses while any consequence card, any event with variants, or any armistice card with an early_text (R-WAR-008) has no reviewed_by (a person's sign-off), alongside the sensitivity gate (Q-ALT-9). Timeline.Entries lists each resolved event, in window order, with the player's option beside the record's, and Timeline.Origin names the road a consequence card follows.

R-ALT-012: The historical run (the reference player's orders from 15 May 1948 to 15 January 1952) raises no consequence card and diverges nowhere.

Gameplay camera (gdd-0.3 only)

R-CAMERA-001: Only exact gdd-0.3 fixes gameplay yaw at zero (north), including entry, input frames and minimap settling. Orbit input has no effect. Legacy and gdd-0.1 retain orbit. Pan, the 30 px drag onset, fixed pitch and view-axis zoom retain their existing behavior, rates and 6 to 250 km distance range. Explicit promo poses remain separate. Heading and zoom: owner ruling, 27 October 2026, #1720. Camp marker and label dimensions remain subject to measurement and owner approval.

Population identity foundation (gdd-0.3 only)

These contracts cover #1689 steps 1–2 and the Historical start inventory (R-POP-004) only. Production era selection, Food consumption and command/save/reload integration remain held; helper coverage does not complete #1689. See population-ledger.md.

R-POP-001: Only exact gdd-0.3 enables the population-ledger capability, independently of the broad GDD loop and Attention (R-LOOP-002/005). An immutable own-population record has a dedicated nonnegative ordinal id. Synthetic fixture inventory is ordered by source key (ordinal string comparison), then within-source ordinal. Registered arrivals allocate monotonically; only explicit combat losses remove living ids. Sources remain registered after loss, and ids are never reused. Reconciliation checks the exact created-minus-lost id set against the living ledger, not just equal totals. Status/region counts include every matching record once. The gdd-0.3 scale of a synthetic or Open fixture reads one positive Agreed integer from population.people_per_unit in its catalogue (currently 10,000, Open mode's scale under D2); a Historical fixture reads rules.yaml's people_per_unit for the start day (1,000), as R-UNIT-006 does (owner, 1 November 2026, D2: "Approve all six (Recommended)", "The Historical MVP keeps main's start: census cities at 1,000 people a unit (R-CITY-007)"; R-POP-004); rows with a census/, units_1948/ or field/ source key are the Historical start's and are refused in any other fixture. Old identities retain the common rules scale. Missing, fractional, nonpositive, overflowing or Deferred values have no fallback, in either source.

R-POP-002: A ledger record has exactly one typed status and assignment, validated location, home and captor references, whole reception-turn ticks capped by the catalogue requirement, civilian profession-specific tier and actual-work XP, separate military qualification/combat XP, and distinct issued civilian/military equipment and paid facts. Unknown references, unresolved teacher ids, incompatible statuses/assignments, inconsistent region/hex/facility bindings, invalid progress/tiers and duplicate issued-item ownership are refused before mutation. Moves and atomic status/location/assignment transitions retain identity/history. Capture retains home context without inventing a prison hex or refunding equipment. Ledger changes never touch resources, causes, RNG or Attention. Initialization is internal and fixture-only. Simulation.Step takes a game whose ledger was initialized with mode Historical, the dark start's, and only that (#1725 PR 6; owner, 8 November 2026, Draft 10, "Historical mode alone (Rec.)"): it runs the turn's orders on its working copy, and an order whose system is not ledger-ready refuses there with that system's guard, so the game is unchanged ("Yes, unready refuses (Rec.)"). It hands over delegated event cards only when one is waiting, and a waiting card is still refused ("Only if none pending (Rec.)"). TurnResolver.Resolve, Simulation.Preview and Simulation.DelegatedTo still refuse any initialized ledger until a separate dark-turn PR ("Separate dark-turn PR (Rec.)"), so a Step on a dark start ends refused by the turn with the game unchanged. A synthetic or Open ledger is refused at Step's entry as before. Saving a dark game waits for #1725 PR 8 ("Not yet, PR 8 (Rec.)", R-POP-003). The dark start (#1725 PR 5) is the internal, fixture-only Simulation.NewGame overload that initializes the Historical ledger (R-POP-004) in place before Pools.Init: Pools.Init then reads the ledger's need, one ration for each living own population id (R-FOOD-001, #2470), after the ledger reconciles, and StartingForces.Init places the forces once each own starting formation's strength is its ledger soldiers, refusing otherwise. Both read only a ledger initialized with mode Historical, the dark start's; a synthetic or Open ledger is refused by both, as by the other aggregate systems, with the state unchanged. An aggregate system whose own rule says it runs on that Historical ledger (#1725 PR 7: civilian upkeep, R-MONEY-002, the first, and food_supply's stock term, R-FOOD-003; training, R-TRAIN-001 to R-TRAIN-004; military upkeep, R-UNIT-010; arms workshops, R-ARMS-005 and R-ARMS-006; Send home, R-UNIT-005) also reads only that ledger; a synthetic or Open ledger is still refused by it. Prisoners.Take and Prisoners.Upkeep are such systems (#1710, R-POW-001, R-POW-002, R-POW-005): a capture turns the surrendered formation's serving records into captives with their captor and home, all or nothing, and records no loss. EventSystem.Trigger still refuses a ledger, so a dark start fires no opening event; the public NewGame is unchanged. The step-7 Food settlement given a month's synthetic context is another exception to that refusal: it reads its need from that month's frozen snapshot (R-FOOD-002, #2470). The live month snapshot (Gdd03SyntheticMonthContext.CreateLive, #1725 PR 7; owner, 8 November 2026, "One snapshot per turn (Recommended)") reads only a ledger initialized with mode Historical and changes no record; a synthetic or Open ledger and an aggregate game are refused (R-FOOD-002). The step-6 harvest given that live month (Pools.Harvest, R-FIELD-002, #1694) also runs on that ledger: it reads only a ledger initialized with mode Historical, a synthetic or Open ledger is still refused by it, and without a month it is refused on any initialized ledger. Old identities refuse ledger mutation. Prices, progression awards, return destinations, tool lock-in and captive-equipment outcomes remain with their owning issues. A craftsman's tier is novice, apprentice or expert (R-CRAFT-003); a farmer's or pioneer's is novice, trained or veteran.

R-POP-003: Clone copies all ledger containers; immutable records and frozen bindings may be shared. The canonical writer emits a gdd-0.3-only section ordered by numeric population id, with fixed field order, stable machine tokens, invariant integer formatting and length-prefixed reference strings. It covers every record/assignment/equipment/reference field, binding, creation/loss provenance and allocator even after the last combat loss. Legacy/gdd-0.1 emit no ledger section or allocator marker. A location's optional city (a regions.yaml member location of its region that is a census city of cities.yaml or one of that region's field hosts, where R-POP-004 houses own people: a settlement hosts only F19 and F20, owner, 7 November 2026, #1689: "These two only (Recommended)"; any other, including a locality, a sites.yaml id or a map location, is refused when a fixture is created, naming the city and the region, R-POP-002) is written after its other fields as |city= with a length-prefixed id, only when set, as R-UNIT-002's |home= is, so a location without a city keeps its canonical text and hash. R-SAVE-001/002 still govern identity mismatch refusal and replay: no ledger snapshot or new command syntax is added in this foundation, and fixture injection is not replay evidence. Existing unpublished schema-4 draft hashes change; no migration promise is made.

R-POP-004: The Historical start inventory under exact gdd-0.3 (#1689; owner, 1 November 2026, D2: "Approve all six (Recommended)", and D4). It is read from the same content as the aggregate start: each census city (R-CITY-006) gives its R-CITY-007 start units as free civilians, and each Israeli starting formation of map/units_1948.yaml (R-UNIT-006) gives its strength as active soldiers. The 20 farmer units are added to the census total (owner, 7 November 2026: "On top, change the rule (Recommended)", which reverses this rule's earlier "no more and no less"), one a field, with source key field/<label> (F01 to F20 of regions.yaml field_hexes) and ordinal 0 (maintainer call, the coordinator, 7 November 2026, so the key stays if a host changes); see the farmers below. Enemy formations are not own population, and a dated arrival (R-UNIT-009) is not in the start inventory: it takes new ids when it arrives (Q11), as R-UNIT-009 records; its service target and region are bound at the start with no record (coordinator call, 8 November 2026). Counts use rules.yaml's people_per_unit (1,000, D2), as R-CITY-007 and R-UNIT-006 do; the gdd-0.3 catalogue's population.people_per_unit (10,000) is Open mode's (D2). A city's records have source key census/<city id> and a formation's units_1948/<row id>, each with within-source ordinals from 0, so R-POP-001's order (source key by ordinal string comparison, then ordinal) gives the ascending population ids that Q10's "N units" commands take. Legacy and gdd-0.1 content refuse it. It reads no RNG and changes no state. It is not yet wired into the public Simulation.NewGame (only the internal dark start of R-POP-002 calls it): an initialized ledger refuses every aggregate system (R-POP-002), so a new gdd-0.3 game keeps the aggregate start until the producers are integrated. A fixture builder (HistoricalPopulation.LedgerFixture, internal, fixture-only) turns the inventory into an initialized Historical ledger with mode Historical (R-POP-001's 1,000 a record), each record in a region of content/regions.yaml (#1691): a census record takes the region whose member_locations hold its city (census/tel_aviv is tel_aviv_jaffa; Jaffa gets no records), and its location names that city, and a formation record takes the region whose member hexes hold the formation's starting anchor hex (owner, 3 November 2026, #1689 open question 1: "Its anchor hex's region (Recommended)"). A source with no region is refused; no region is invented. Census records are free civilians, with no assignment (but for each city's employed, below) and their reception done: they start at reception.turns, the "actual reception turns before free-civilian status" (docs/gdd/balance.md:91), the ledger's cap (R-POP-002), since R-VET-001 gives work XP only to a received free civilian (an agent reading of balance.md:91, not an owner ruling); formation records keep 0, which no ruling sets. Formation records are active soldiers whose service target is their own formation. Each formation record carries one issued military item, id units_1948/<row id>/<ordinal>, owned and recorded as paid; census records carry none. One census record is the training centre's veteran pioneer teacher (owner, 7 November 2026, #1689: "Yes to all three (Recommended)", whose third part reads "the training centre's one veteran pioneer teacher is a Haifa census unit"): content/gdd03-balance.yaml's start.pioneer_teachers (1, Agreed, "included in population total") gives the count, so no record is added, and the fixture refuses any other value, since no ruling places a second teacher. It is the last census/haifa ordinal (maintainer call, the coordinator eyalgolan-30, 7 November 2026), a free civilian pioneer at the veteran tier (R-TRAIN-005: "a veteran of the matching profession"), with work XP 0 (owner, 7 November 2026, relayed by the coordinator: "0 (Recommended)"), and no assignment, since R-TRAIN-008 refuses a teaching teacher with no pupil; otherwise it is like any other census record. It is "Takeable, but last (Recommended)" (owner, 7 November 2026, relayed by the coordinator): there is no guard, and as the highest census/haifa id it is the last record that Q10's "N units" commands and #1692's job default reach. No ledger producer runs the job default or a Train order yet. Issued items are apart from the spare Arms: the fixture leaves the Arms stock untouched, and a new game still starts with arms_start's 12 (#1717, 27M-24: "The starting formations are equipped, and the Arms stock starts at 12"). Each city's employed hold Industry work (#1725 PR 3; owner, 7 November 2026, #1689: "Start in Industry (Recommended)", so every employed unit starts in Industry; the count is R-CITY-007's people x share, "Keep people x share (Recommended)"). Each census city has one Industry Work target, industry/<city id>, in its region, naming the city, with the city's job_slots as its capacity (R-CITY-007); start.industry_slots (8) is the 10,000-scale value and is never a city's capacity. The city's first employed census ordinals, its ascending population ids, hold Work on it, as #1692's job default gains the Industry assignment (owner, 3 November 2026: "Approve all three (Recommended)"); the rest have no assignment, so the teacher, the last ordinal, is not among them. The ledger refuses a record on an Industry target of another city (Q27: "a worker works in their own city") and more records than the capacity (R-POP-002). The live month snapshot has one entry per Industry target, available with its capacity, as a field's (owner, 11 November 2026, #1725: "A: available now (Recommended)"). The aggregate employed stays as it was and equals the ledger's Industry count at the start; capacity changes after the start (projects, events) and the job default on the ledger come later. The opening field farmers are (#1725 PR 4; the owner's 1:2:1 novice/trained/veteran ratio applies per host, which gives 1 / 18 / 1 nationally: owner, 7 November 2026, "Per host, 1/18/1 (Recommended)", #1689). HistoricalPopulation.FieldFarmerHosts (internal) gives each regions.yaml field_hexes host its fields' farmers, one a field (owner, 3 November 2026, #1694: "One farmer per field (Recommended)"), split 1:2:1 per host by Gdd03FarmerTierSplit in field-label order; the 20 match start.prepared_fields (Agreed), and any other total, or a field with no host, is refused. Each farmer is a free civilian farmer staffed on its field from turn 1 (owner, 7 November 2026, #1689: "Yes to all three (Recommended)"): its assignment is Work on the field's target (field/<q>,<r>, as Gdd03FieldFixture), its location the field's region with its host as the city (Nir Am and Nirim for F19 and F20, R-POP-003). Tiberias hosts six fields, F03 novice, F04 to F07 trained and F08 veteran (owner, 7 November 2026, relayed by the coordinator: "F08 veteran, F03 novice (Recommended)"); every other host has one or two, all trained. Work XP is 0 for every tier (owner, relayed by the coordinator: "0 for all (Recommended)"; veterans "0 (Recommended)", #1689). Reception is done (reception.turns), as for the census (maintainer call, the coordinator, 7 November 2026). No tools: post-MVP (#1693, owner, 3 November 2026, answer 1: "Tools in the MVP: no."). field/ rows, like census/ and units_1948/ rows, are refused outside a Historical ledger (R-POP-001). Farmers hold a field, not a job, so they are not among a city's employed (R-CITY-007) or civilians, not in R-OBJ-004's sums, and not among the people without a job a Train order takes (R-TRAIN-001): the dark start's aggregate cities keep the census's counts. The dark start adds one housing place per farmer to each census host's housing capacity (same ruling); Nir Am and Nirim have no cities.yaml record and get none for now (owner, relayed by the coordinator: "None for those two for now (Recommended)"), so 18 of the 20 places exist until the housing producer (#1725 PR 7). The public NewGame is unchanged. The formations' rows 5 to 8 are in the fixture (owner, 7 November 2026, #1689 comment 6704389240: "Veteran for all 8 (Recommended)" and "1 each, Negev 0 (Recommended)", Negev's 0 a labelled stand-in). The veteran level lives on the formation, a veteran_soldier (R-UNIT-006) that is each soldier's ActiveService target, so a formation record's military qualification is its type's unit_types.yaml class (infantry), the source basic-training graduates use, with no new level field (maintainer call, eyalgolan-30's coordinator, 7 November 2026). Its combat XP is content/gdd03-start.yaml's formation_combat_xp: 1 for Golani, Carmeli, Alexandroni, Kiryati, Givati, Etzioni and Harel, 0 for Negev. The list also names each own brigade dated after the start (R-UNIT-009): 1 for the 7th Brigade (owner, 8 November 2026: "1, like the others"). The loader requires the list, refuses a row that is not an own opening or dated formation, a duplicate, a missing formation or a value outside 0 to military.combat_xp_full, and the fixture refuses without it (no fallback). Only the ledger fixture reads it: the units NewGame starts with keep 0 combat XP (same maintainer call); a veteran_soldier gains none (R-UNIT-011). The paragraphs below record the owner's decision on each. The ledger scale for this start is no longer Deferred: a Historical ledger fixture counts each record at rules.yaml's people_per_unit for the start day, 1,000 people (R-POP-001; D2: "The Historical MVP keeps main's start: census cities at 1,000 people a unit (R-CITY-007)"). The four opening-ledger values were Deferred pending a separate sourced draft for owner approval (4 November 2026, Q24; the coordinator asked: "How should the four deferred opening-ledger values be approved before Historical NewGame initializes the ledger?"; the owner answered: "do the recommended for all"; the chosen option read: "Request and approve a separate sourced draft for employed assignments, farmer identities, formation qualifications and combat XP before wiring NewGame"; recorded on #1689). Farmer home cities needed a source-backed city-by-field proposal for owner approval (4 November 2026, Q26; the coordinator asked: "How should the census-city home of each opening field’s farmer be selected when its region contains several cities?"; the owner answered: "do the recommended for all"; the chosen option read: "Have an agent prepare a source-backed city-by-field proposal for owner approval"; recorded on #1717).

Owner decision on both drafts, 7 November 2026: "all reccomended in "Decisions for the owner: the GDD 0.3 starting population (#1689) and where the field farmers live (#1717)" - approved" (recorded on #1689). This wires nothing into Simulation.NewGame.

Approved, starting population (research_notes/GDD phase 4/opening-ledger-values-1689.md, Owner-approval table). Row 1: employed units = census units x the printed working share, for Jerusalem, Tel Aviv-Yafo, Haifa, Tiberias and Safed (the wording answer below keeps R-CITY-007's people x share; those five cities' rounded units are the same either way). Row 2: the other seven census cities use the national 43.9%, labelled a stand-in. Row 4: no farmer identities for now. Row 5: one training level for all six Field Corps brigades, the level being a labelled stand-in. Row 8: Negev's combat experience is a labelled stand-in. The answers below give both values.

Approved, field hosts (research_notes/GDD phase 4/opening-field-farmer-cities-1717.md). Every row settled by a source, and every row forced by the same-region rule. F01 and F02: Rosh Pinna. F11: Hadera. F18: Rishon LeZion. Decision A, option 1: F19 and F20 keep the same-region rule and are hosted at Nir Am and Nirim.

Owner answers, 7 November 2026, on the items that decision left open (recorded on #1689). Farmers: "On top of census (Recommended)": the 20 farmers are separate units in their host towns, not taken from the city population. Rows 5 and 6: "Veteran for all 8 (Recommended)". Rows 7 and 8: "1 each, Negev 0 (Recommended)"; Negev's 0 is a labelled stand-in. Decision A host kind: "These two only (Recommended)": a settlement can host farmers for F19 and F20 only. F13, F14, F17: "Haifa, Haifa, Rehovot (Recommended)", design choices, not sourced. Nirim: "Add Nirim to Negev (Recommended)". Row 1 wording: "Keep people x share (Recommended)", as R-CITY-007 and the code do.

Row 3, the mapping from economic branches to game jobs: the owner first answered "Map branches to game jobs", then (recorded on #1689) "Start in Industry (Recommended)" for building workers and "Start in Industry (Recommended)" for city people in agriculture. Trade, transport and services have no game job, so they go to Industry too (design choice). Row 3 closes with no change in the game: every employed unit starts in Industry. The printed branch shares stay in the research note.

Owner answers, 7 November 2026, on the start's population rules (recorded on #1689). "On top, change the rule (Recommended)": the 20 farmer units are added to the census total. "Yes to all three (Recommended)": every farmer is staffed on their field from turn 1; one added housing place per farmer in their host; the training centre's one veteran pioneer teacher is a Haifa census unit.

No item of either draft is open. These values are not yet wired into Simulation.NewGame: a new gdd-0.3 game keeps the aggregate start.

The opening field hosts are data under exact gdd-0.3 (#1717): regions.yaml's field_hexes rows carry field, host, host_basis and host_source, and PlayerRegionGeography.FieldHosts reads them. F01 to F20 are the fields' places in the file's order, as in the research note, and the loader refuses any other label. Hosts: F01 and F02 Rosh Pinna; F03 to F08 Tiberias; F09 and F11 Hadera; F10 and F12 Zikhron Ya'akov; F13 and F14 Haifa; F15 and F16 Petah Tikva; F17 Rehovot; F18 Rishon LeZion; F19 Nir Am; F20 Nirim. host_basis keeps the note's evidence apart: sourced (F09, F10, F12, F15, F16), administrative (F07, F08), forced (F03 to F06), inference (F01, F02, F11, F18), design_choice (F13, F14, F17, the owner's choice, not sourced) and decision_a (F19, F20). The loader refuses a settlement host on any field but F19 and F20, a census-city host outside its field's region, a settlement that is not a member of the field's region, an unknown basis and a host with no host_source, naming the file and field. Nirim is a member location of the Negev; the legacy and gdd-0.1 identities read none of this. Nothing else reads the hosts: there is no farmer record and no NewGame wiring.

#1706 slices 0/1 (coordinator decision, 26 Oct 2026, Q5): the synthetic combat catalogue foundation is inactive in the actual game. R-COMBAT-001 through R-COMBAT-008 retain their current runtime meanings. See the catalogue foundation for its data contract and sources.

Map editor infrastructure (gdd-0.3 authoring only)

R-MEDIT-001: The versioned authoring envelope changes override fields only. It carries no names, source/review metadata, population, jobs, capacities or campaign state. Region, field and facility ids in synthetic tests are fixture-only. No production integration or historical data is authorized by this infrastructure slice.

R-MEDIT-002: Only an explicit gdd-0.3 editor context accepts the versioned envelope. Version, ruleset, draft flag, base revision/digest and all edit lists are required; unknown properties, incomplete axial coordinates and duplicate edit identities are refused. Base revision and registry references are single-line tokens without whitespace or control characters. Serialization is invariant and ordered without mutating the candidate. The authoring envelope uses a separate schema; the legacy schema stays byte-identical. Legacy parsing/serialization, content, save/state bytes and goldens remain unchanged. Failed document replacement retains both the current candidate and rejected raw input; serialized source preserves edits and provenance through Undo and domain reload.

R-MEDIT-003: Schema validity is not semantic validation. Given the loaded content, corridor, force-anchor, target and objective bindings are checked against the map, with the overlay's location moves applied to targets and objective sites, and a broken one is a blocking MAPEDIT_DEPENDENCY_BROKEN row naming its record (Q58). Without content, those four stay unavailable. Missing region/field/facility adapters and an unapproved base-manifest contract are named blocking diagnostics, even for an empty candidate. Base mismatch reports both identities. This infrastructure cannot canonical-save or export, and has no runtime reader or disk writer. Integrating domain validators, persistence and the headed editor is separate work.

R-MEDIT-004: A visual or label offset (#1721 slice B; owner, 1 November 2026, Q58 and Q59) is one offsets row per location id and kind (visual or label): the axial hex (q, r) plus whole metres east and north of that hex's centre. It is valid while the point stays inside its own hex, the map's 4.4 km pointy-top hex (hex_size_km in content/map/hexes.json is the circumradius; the boundary counts as inside). A point outside it gives a MAPEDIT_OFFSET_OUTSIDE_HEX warning that names the row (offsets[<id>/<kind>]) and is not blocking; every other diagnostic stays blocking. The point is kept as written: no offset, location or sourced geometry is moved to clear the warning. Offsets serialize ordered by id and kind, so save and reload give the same bytes.